The Geography of Equity Analysis

Christopher J. Malloy∗

October 27, 2003

ABSTRACT

I provide evidence that geographically proximate analysts are more accurate than other an-

alysts. Stock returns immediately surrounding forecast revisions suggest that local analysts

impact prices more than other analysts. These effects are strongest for firms located in small

cities and remote areas. Collectively these results suggest that geographically proximate ana-

lysts possess an information advantage over other analysts, and that this advantage translates

into better performance. The well-documented underwriter affiliation bias in stock recom-

mendations is concentrated among distant affiliated analysts; recommendations by local af-

filiated analysts are unbiased. This finding reveals a geographic component to the agency

problems in the industry.

∗London Business School. I thank Nick Barberis, Karl Diether, Gene Fama, Gilles Hilary, Rick Green, SteveKaplan, Owen Lamont, Richard Leftwich, Lior Menzly, Toby Moskowitz, Lubos Pastor, Raghu Rajan, an anony-mous referee, and seminar participants at London Business School, Ohio State, University of Chicago, University ofVirginia, Yale, and the AEA annual meetings for helpful comments and suggestions. I also gratefully acknowledgethe contribution of Thomson Financial for providing earnings per share forecast data, available through I/B/E/S. Thisdata has been provided as part of a broad academic program to encourage earnings expectations research. Pleasesend correspondence to: Christopher Malloy, London Business School, Sussex Place, Regent’s Park, London NW14SA, UK, phone: 44-207-262-5050 x3278, email: cmalloy@london.edu.

The Geography of Equity Analysis

ABSTRACT

I provide evidence that geographically proximate analysts are more accurate than other an-

alysts. Stock returns immediately surrounding forecast revisions suggest that local analysts

impact prices more than other analysts. These effects are strongest for firms located in small

cities and remote areas. Collectively these results suggest that geographically proximate ana-

lysts possess an information advantage over other analysts, and that this advantage translates

into better performance. The well-documented underwriter affiliation bias in stock recom-

mendations is concentrated among distant affiliated analysts; recommendations by local af-

filiated analysts are unbiased. This finding reveals a geographic component to the agency

problems in the industry.

The Geography of Equity Analysis

This paper investigates the effect of distance on the accuracy and investment value of equity

analysts’ forecasts and recommendations. Using a large panel of analyst data from 1994-2001,

I provide evidence that geographically proximate analysts outperform their distant counterparts.

Specifically, local analysts are significantly more accurate than other analysts. I find this effect to

be strongest in small firms, firms located outside of the most populated cities, and firms located in

remote areas. The magnitude of the accuracy advantage is small in the overall sample ($0.025 per

share on average), but larger in various subsamples; for example, local analysts covering small

stocks in remote areas are approximately $0.141 per share more accurate than distant analysts.

Abnormal returns surrounding large forecast revisions suggest that local analysts also impact

stock prices more than other analysts. In tests controlling for the magnitude of the revision,

firm size, and analyst affiliation status, I find that local analyst revisions are associated with

an incremental average excess return of -0.128% per day (t=-2.46) in the three days surrounding

strong negative revisions; this number grows to -0.164% (t=-2.54) for firms located in small cities.

Following strong positive revisions, these figures are slightly smaller (0.095% for all stocks, and

0.102% for non-metro stocks), but still significant. As with the accuracy results, these effects are

strongest in the most recent subperiod.

These findings are consistent with the hypothesis that geographically proximate analysts pos-

sess an information advantage over other analysts. As an information story would suggest, local

analysts appear to have a decided advantage covering small stocks, and stocks located in remote

areas, where the access to private information is likely to be strongest and competition for that

information the weakest. By contrast, I find little support for the notion that the link between

geography and performance is a natural by-product of the endogenous coverage decisions made

by analysts. For example, local analysts do not outperform other analysts simply as a result of

covering fewer stocks or through increased specialization. I present some evidence that local

investor demand or visibility may be linked to local abnormal performance, perhaps through in-

creased analyst attention or effort, but these results are based on a small subset of stocks for

which advertising expense is nonzero.

1

Geography also provides an interesting viewpoint through which to explore the much-publicized

affiliation effect. The affiliation effect refers to the perceived tendency of analysts from a broker-

age house that has an underwriting relationship with a stock to issue more positive predictions

than analysts from non-affiliated houses. The implication is that brokerage houses reward opti-

mistic affiliated analysts who generate investment banking business and trading commissions, at

the expense of clients who trust analysts’ research to be unbiased.1 The extent to which affiliated

analysts’ research suffers from this potential agency problem is an open question, and one that

has spawned a series of recent empirical papers.2 Since I am able to separate out the effects of

affiliation and geographic proximity throughout the paper, another contribution of this paper is

that I am able to quantify both agency costs and information costs for a large sample of analysts.

In this context, I explore the effect of distance on the well-documented affiliation bias in an-

alysts’ stock recommendations. Lin and McNichols (1998) examine the effect of underwriting

relationships on stock recommendations, and report that lead underwriter analysts’ recommenda-

tions are significantly more favorable than those made by unaffiliated analysts. They also report

that three-day returns to lead underwriter “Hold” recommendations are significantly more nega-

tive than those to unaffiliated “Hold” recommendations.3 These results are typically interpreted

as signalling an agency problem on the part of equity analysts, specifically that analysts face a

conflict of interest between providing high quality, unbiased information to their clients, while

at the same time aiding the corporate finance wing of their brokerage firm. I re-examine these

results by conditioning on the location of the analyst.

I am able to replicate the underwriter affiliation bias in stock recommendations in my sample,

but find no evidence of a bias for affiliated analysts that are also located nearby the firm being

covered. For example, the incremental three-day average excess return to distant affiliated “Hold”

recommendations is -1.74% (t=-3.87), while the incremental market reaction to local affiliated

“Hold” recommendations is only -0.25% (t=-0.78). I argue that this is because local affiliates

1See Hong and Kubik (2003) for evidence in support of this view.2See, for example, Lin and McNichols (1998), Dechow, Hutton, and Sloan (1998), and Michaely and Womack

(1999).3Additionally, Michaely and Womack (1999) report that stocks for which underwriter analysts issue “Buy” rec-

ommendations perform more poorly than “Buy” recommendations by unaffiliated analysts prior to, at the time of,and subsequent to the recommendation date. By contrast, Dugar and Nathan (1995) find no evidence of stock priceresponse to recommendations by affiliated or unaffiliated analysts.

2

are less likely to be working at high-status firms, where agency problems are intense due to the

constant pressure to garner investment banking business. Thus, not only does this paper point

to an informational link between proximity and equity analysis, but it also sheds light on the

geographic nature of the agency problems in the industry.

The paper is organized as follows. Section I provides some background and a description of

the data, while Sections II-IV present my empirical design and results. Section V concludes.

I. Methodology

A. Background and Motivation

Arguments supporting a link between proximity and information flow are presented in Coval and

Moskowitz (2001), who analyze the role of geography in the context of mutual fund managers.

They maintain that geographic proximity is inversely related to the cost of information acquisi-

tion. It also provides access to private information: “Investors located near a firm can visit the

firm’s operations, talk to suppliers and employees, as well as assess the local market conditions

in which the firm operates.” Similarly, I would argue that the ability of local analysts to make

house calls rather than conference calls, during which time they can meet CEOs face-to-face and

survey the firm’s operations directly, provides them with an opportunity to obtain valuable pri-

vate information. Following this logic, geographic proximity is a sensible proxy for the quality

of analyst information.

On a more basic level, the advantage of focusing on equity analysts is that this industry

offers an ideal testing ground for a number of theories of economic behavior. Since analyst

data is available in large quantities and relatively standardized formats, this industry is one of

the few areas that allows precise estimation of the effects of asymmetric information, agency

costs, herding, etc., for an important segment of the financial community. I use this testing

ground to evaluate the idea that geographic proximity facilitates information flow. Further, since

I can control for investment banking affiliations, I attempt to distinguish information effects from

agency effects throughout the paper.

3

Evidence that affiliated analysts suffer from an agency problem caused by the inherent con-

flict of interest in the functions they perform is mounting (see Hong and Kubik (2003) for a

summary). Although an analyst could conceivably gain inside information about a firm that has

an underwriting relationship with the analyst’s brokerage house, and hence deliver more accurate

forecasts or recommendations, previous research (see, for example, Lin and McNichols (1998))

finds that this is usually not the case. This perceived agency problem is garnering widespread

attention in the political arena as well. Congressional hearings to “analyze the analysts” have

already taken place, and on May 8, 2002, the SEC approved proposed rule changes by the NASD

and the NYSE to address security analyst conflicts of interest.

Another policy question that indirectly motivates my research is the recent passage of Regu-

lation FD. Effective October 23, 2000, companies must reveal any “material” information to all

investors and analysts simultaneously in the case of intentional disclosures, or within 24 hours

in the case of unintentional disclosures. According to SEC Proposed Rule: S7-31-99, regulators

believe that allowing selective disclosure is “not in the best interests of investors or the securities

markets generally.” And yet, empirical evidence on whether this type of disclosure even exists in

the first place, or if the enactment of the law has actually reduced selective disclosure, is sparse.4

Since my tests explore a specific possible channel of selective disclosure, they are relevant to this

debate.

My paper also contributes to the growing literature on the importance of geography in eco-

nomics. Coval and Moskowitz (1999), Zhu (2002), and Huberman (2001) report strong pref-

erences for geographically local equities among investors. Hong, Kubik, and Stein (2002) link

geographic proximity and mutual fund managers as in Coval and Moskowitz (2001), but instead

focus on the information that investors that are close together can pass on to one another; they

document a word-of-mouth effect whereby local mutual fund managers are more likely to hold a

particular stock if other managers from different fund families in the same city are holding that

same stock. Similarly, Grinblatt and Keloharju (2001) report that investors are more likely to

buy, hold, and sell stocks of Finnish firms that are located close to the investor, while Portes and

Rey (2002) find a strong geographic component in cross-border equity flows. Petersen and Rajan

4See Agrawal and Chadha (2002) and Bailey, Li, Mao, and Zhong (2003) for a discussion of these and relatedissues.

4

(2002) and Berger, Miller, Petersen, Rajan, and Stein (2002) also explore the concept of physical

distance, but do so in the context of commercial banks’ lending to small companies.

And yet, only a few papers explore geography in the context of equity analysis. In addition,

all of the existing studies focus on cross-border effects, rather than within-country effects. The

international evidence appears to be mixed. For example, Chang (2002) explores the Taiwanese

market and finds that expatriate analysts located outside the country (but whose firms have a local

research group) outperform otherwise similar foreign analysts, but also outperform local analysts.

Similarly, Bacmann and Bolliger (2001) find that foreign financial analysts outperform home-

country analysts in Latin American emerging markets. On the other hand, Bolliger (2001) reports

an accuracy advantage for home-country analysts at small and medium-size brokerage houses

in Europe, and Orpurt (2002) finds that home-country analysts covering German-headquartered

firms forecast earnings better in the short-run than foreign analysts. However, none of these

papers tests a pure distance effect, since borders introduce a host of other issues. As such, by

focusing solely on U.S. equity analysts covering domestic firms, this paper is the first to isolate

the effect of physical distance on the accuracy and investment value of analysts’ forecasts and

recommendations.

B. Data

My sample merges several datasets, the most important of which is the I/B/E/S analyst forecast

data. I use the I/B/E/S U.S. Detail History and the I/B/E/S Recommendation History datasets for

this paper. The Detail History file contains individual analysts’ earnings forecasts of U.S. compa-

nies between 1983 and 2002, while the Recommendations History file contains analysts’ invest-

ment recommendations (translated by I/B/E/S into a common 1-5, 1=“Strong Buy,” 5=“Strong

Sell” scale) issued from 1993 to 2002.5

5I use the I/B/E/S adjusted earnings-per-share data in this paper, since I require a smooth time-series of earningsforecasts. This opens up the possibility that my results are affected by the adjustment bias discussed in Diether,Malloy, and Scherbina (2002) and Payne and Thomas (2002). However, I can report that my results are slightlystronger when I use the raw, unadjusted data.

5

In particular, since I match each analyst with his firm, and then match each firm to its ge-

ographic location, I also use the Broker Translation File. Using the Broker Translation File, I

match the analyst codes from the Detail File to the analyst’s name and brokerage firm.6 I then

match the analyst’s name and brokerage firm name (by hand) with entries from Nelson’s Direc-

tory of Investment Research. This volume is available annually from 1994-2002, and contains

each analyst’s name, her phone number, which office she works in, and the address (city, state,

and zip code) for each of the firm’s offices (including all branch offices). Therefore I can distin-

guish between analysts who are located at the brokerage firm’s headquarters and those who are

not. Each volume of Nelson’s Directory of Investment Research is published in December of year

t using data as of November of yeart. I classify an analyst’s location starting in November of

yeart and lasting until October of yeart +1 according to the information in Nelson’s Directory.7

I then find latitude and longitude data for each analyst’s location (measured at the city centroid)

using the U.S. Census Bureau’s Gazetteer city-state (places.zip) file.

I obtain the location of each firm’s headquarters each year from the Compact Disclosure

database, and find the latitude and longitude data for each firm’s location using the U.S. Census

Bureau’s Gazetteer city-state file as before. Using the latitude and longitude data, I then compute

the distance between each analyst and each firm that she covers.

In order to control for brokerage house affiliations, I use data from the SDC New Issues

Database. Using SDC, I obtain information on initial public offerings (IPOs) and secondary eq-

uity offerings (SEOs) conducted during my sample period, including the date of the offering and

the name of the lead underwriter. I merge information on these IPOs and SEOs with my ana-

lyst sample, so that I can categorize for each stock that an analyst covers whether her brokerage

house has an underwriting relationship with that stock. More specifically, if an analyst issues an

earnings forecast on a stock within a specified period (see below) after its IPO/SEO date and in

which her brokerage house is the lead underwriter for the IPO/SEO, then I define that stock as

having an underwriting relationship with the analyst’s brokerage house.

6I employ a similar procedure for the Recommendations History dataset. I also eliminate all analyst teams fromthe dataset, since it is not always possible to match them up to a unique location.

7My results are not sensitive to this timing convention. For example, if I classify an analyst’s location starting inNovember of yeart and lasting until October of yeart +1 only if the analyst appears again in the Nelson’s Directoryin November of yeart +1, my results are virtually identical.

6

Stock returns are drawn from the Center for Research in Securities Prices (CRSP) Daily

Stocks Combined File, which includes NYSE, AMEX, and Nasdaq stocks. For some of the tests

in the paper, firms must also have data in COMPUSTAT on book equity for the fiscal year ending

in calendar yeart−1. Since I only have data from the Nelson’s Directory of Investment Research

starting in November 1994, my main sample period extends from November 1994 to December

2001 in order to avoid a survivor bias that might be caused by applying locations backward to

1983 (the first year for which the I/B/E/S Detail data is available).

I also collect data on individual analysts’ reputations by using the list of All-Americans drawn

from the October issue ofInstitutional Investormagazine for 1994-2002. I classify an analyst

as an All-American starting in November of yeart and lasting until October of yeart +1 if she

appears in the list of All-Americans in October of yeart. Although I collect data for First, Second,

and Third Team All-Americans, as well as Honorable Mention All-Americans, the tests in this

paper use the term All-American to refer to an analyst listed in any of these four categories. As in

Hong and Kubik (2003), my primary measure of the brokerage house hierarchy is derived from a

brokerage house ranking published byInstitutional Investormagazine. Each year in the October

issue, the ten or so brokerage houses with the most All-Americans are listed as “The Leaders.” I

classify the top ten houses in this annual poll as high-status starting in November of yeart and

lasting until October of yeart +1 if the house appears in the list of “The Leaders” in October of

yeart; all other brokerage houses not on the list are classified as low-status using the same timing

convention.

Table I provides descriptive statistics of my final sample, which is formed by merging these

various datasets. Panel A presents the summary statistics for the one-year earnings forecasts

sample, broken down by forecast type and firm type; Panel B presents the same statistics for

the recommendations sample, broken down by recommendation type and firm type. The entire

one-year earnings forecasts sample includes 4344 firms located in all 48 states of the continental

U.S. (firms and analysts located in Alaska and Hawaii are excluded from the sample so as not

to exaggerate the effect of physical distance), plus the District of Columbia (classified as a state

in Table I). The sample also includes 4254 analysts located in 34 different states. While a large

number of forecasts are issued by analysts located within 100 kilometers of New York City (Pct.

7

NYC=56.47%), my results are not affected when these observations are removed from the sample

(see Table VI). Panel B shows a similar distribution of coverage for the stock recommendations

sample.

Similar to Coval and Moskowitz (2001), I place firms, forecasts, and recommendations into

three categories: “Metro” - defined as being located in any of the 20 most populated cities; “Non-

Metro” - defined as not being located in any of the 20 most populated cities; and “Remote” -

defined as being at least 400 kilometers away from any of the 20 most populated cities. The 20

most populated cities are defined by the U.S. Census Bureau at the beginning of each year. Not

surprisingly, analysts in remote and non-metro areas tend to be associated with smaller brokerage

firms (14.88 analysts on average for remote analysts compared to 42.58 analysts on average for

metro analysts). Similarly, remote firms tend to be smaller, with less analyst coverage. The

recommendations sample shows similar patterns, although the analysts tend to work at slightly

larger firms, and firms tend to have fewer outstanding recommendations.

A subtle shift also occurs between the two subperiods, as the percentage of forecasts issued

by analysts located within 100 kilometers of New York City drops from 59.08% to 54.55%. This

shift is not present in the recommendations sample, however. In addition, untabulated statistics

indicate that in the early period (from November 1994 to December 1997), the percentage of

forecasts issued by analysts within 100 kilometers of San Francisco was 6.44 percent of the

total sample; by contrast, in the period from January 1998 to December 2001 this percentage

jumped to 9.25%. A similar shift (6.51% of recommendations in the early period up to 9.41% of

recommendations in the later period) shows up in the recommendations sample.

II. Relative Forecast Accuracy

A. Regression Methodology

I run cross-sectional regressions of relative forecast accuracy on a host of analyst characteristics

in order to test the notion that geographically proximate analysts possess superior information.

My framework is analogous to the one developed in Clement (1999), in that I control for firm-year

8

variation in both the dependent and independent variables. Clement (1998) finds that controlling

for firm-year effects increases the likelihood of identifying systematic differences in analysts’

forecast accuracy relative to a model that controls for firm fixed effects and year fixed effects.

Firm-year effects result from factors that make a particular firm’s earnings easier (or harder) to

predict in some years than others. Examples of events that may give rise to firm-year effects are

voluntary management disclosures, mergers, strikes, etc.

I measure performance by comparing an analyst’s absolute forecast error to the average ab-

solute forecast error of other analysts following the same stock during the same time period.

Specifically, I calculate the de-meaned absolute forecast error (DAFEi, j,t) which is equal to the

absolute forecast error for analysti’s forecast of firmj for fiscal yeart (AFEi, j,t) minus the mean

absolute forecast error for firmj for fiscal yeart (AFEj,t). Absolute forecast error is equal to

the absolute value of an analyst’s latest forecast, minus actual company earnings (drawn from the

I/B/E/S Actuals File), as a percentage of stock price 12 months prior to the beginning of the fiscal

year. Negative values ofDAFE represent better than average performance while positive values

of DAFE represent worse than average performance. Following Clement (1999), I also employ

an alternate measure of performance,PMAFE (proportional mean absolute forecast error), which

is equal toDAFE divided byAFE; Clement (1998) shows that deflatingDAFE by AFE reduces

heteroscedasticity.

I regress these performance measures (DAFE or PMAFE) on a variety of analyst character-

istics. The model’s independent variables are also adjusted by their related firm-year means to

properly control for firm-year effects; the approach here is equivalent to estimating a model using

firm-year dummy variables to control for firm-year effects. The specific characteristic of interest

for the purposes of this paper is (the log of one plus) the physical distance (LOGD) between

analysti and the headquarters of each firmj she covers.8 In addition to distance, I employ other

measures of geographic proximity. For example, I compute a dummy variable (LOCAL) equal

to 1 if the analyst is located within 100 kilometers of firmj, and zero otherwise. This distance

threshold is arbitrary; however, I have experimented with various distance thresholds, from 50

kilometers up to 200 kilometers, and the results are similar to those presented here. In addition,

8Since locations are identified by latitude and longitude, I calculate the arclength,di j , between each pair. SeeCoval and Moskowitz (1999) for details.

9

my conclusions are unchanged if I replace theLOCALvariable with a dummy variable equal to

1 if the analyst is located in the same state as firmj.

I control for several factors that previous research has identified as contributing to differences

in relative forecast accuracy among analysts. Perhaps the most important of these factors is the

age of the forecast. Clement (1999) reports that relative absolute forecast errors increase at the

rate of 0.35% per day, and stresses the need for careful controls for age when comparing forecasts.

As such, I include (the log of one plus) the number of days (AGE) between analysti’s forecast for

firm j and the firm’s announcement of actual earnings. Following Clement (1999), I also include

controls for experience and available resources. I measure firm-specific experience (FEXP) as

(the log of one plus) the number of days that analysti has supplied a forecast for firmj.9 I measure

available resources by calculating the size of analysti’s brokerage firm (BSIZE), computed as (the

log of one plus) the number of analysts working for the same firm asi and supplying forecasts

during the same fiscal yeart asi. I control for analyst and brokerage firm reputation by computing

dummy variables that are equal to 1 if analysti is an All-Star (STAR) or works for a high-status

brokerage firm (HIGH).10 I also control for the possibility that some analysts are systematically

more optimistic than other analysts by computing the percent of companies (PCTOP) followed

by analysti at timet for which analysti’s last forecast is above actual earnings.

Although Lin and McNichols (1998) report that one-year earnings forecasts are unaffected

by underwriting affiliations, I include a variable (AFFIL) designed to capture such an effect.

Specifically, if an analyst issues an earnings forecast on a stock within five years after its IPO

date or within two years after an SEO date for which her brokerage house is the lead underwriter

for the IPO or SEO, then I define that stock as having an underwriting relationship with the

analyst’s brokerage house. Analysts may deliver upwardly biased estimates for stocks for which

their brokerage house serves as the lead underwriter on a stock offering, since analysts’ pay is

sometimes tied to investment banking revenues. In this case, the variableAFFIL would capture

the effect of this agency problem.

9I have also experimented with a measure of general in-sample experience rather than firm-specific experience,but the results are similar to those presented here.

10See Stickel (1992), Lamont (2002), and Chevalier and Ellison (1999) for insight into the importance of control-ling for these factors.

10

Therefore, the final model takes the form:

DAFEi, j,t ( or PMAFEi, j,t) = β1DAGEi, j,t +β2DFEXPi, j,t +β3DSTARi, j,t +β4DHIGHi, j,t

+β5DBSIZEi, j,t +β6DPCTOPi, j,t +β7DAFFILi, j,t

+(β8DLOCALi, j,t or β8DLOGDi, j,t)+ εi, j,t , (1)

where all variables are firm-year mean adjusted (the D preceding each variable stands for dif-

ferenced or de-meaned). Note that while my test is stated in terms of forecast accuracy, the

regression above analyzes analysts’ forecast errors. Small forecast errors indicate a high level

of accuracy. In addition, the regression equation does not require an intercept since means are

subtracted from all variables.

B. Results

i. Full-Sample Regressions

I estimate Fama and MacBeth (1973) cross-sectional regressions yearly and monthly, and report

the time-series average of the estimated coefficients in Table II. My conclusions are unchanged

if I estimate pooled regressions instead. I restrict the sample to only those firms with at least

three forecasts and two unique analysts in a fiscal year, and I winsorize the extreme 2 percent

of observations for theDAFE, PMAFE, andOPT variables.11 Once again, neither of these

restrictions significantly affects my results.

The estimated coefficients in Table II indicate that a local presence is negatively related to

relative forecast error. Panel A shows that the average yearly coefficient onDLOCAL is negative

and significant in both theDAFE andPMAFE specifications (t=-2.13 andt=-2.28). Meanwhile,

the coefficients onDLOGD are positive for both specifications and significant for thePMAFE

specification (t=2.72). Geographically proximate analysts by either distance metric are thus sig-

nificantly more accurate (i.e., have lower forecast error) in their one-year earnings forecasts than

11I winsorize by applying the 1st and 99th percentile breakpoints of the designated variable each month to obser-vations falling outside these breakpoints in a given month. Using yearly breakpoints or one set of breakpoints forthe entire sample changes no conclusions.

11

their distant counterparts. This result supports the notion that geographic proximity facilitates

information flow, and that distance is a plausible proxy for asymmetric information between an-

alysts. The magnitude of the accuracy effect is small, however; multiplying the coefficient on

LOCAL(=-2.77%) in thePMAFE specification by the mean absolute forecast error for the sam-

ple (=$0.890) reveals a $.025 per share accuracy advantage on average for local analysts.

This main result is robust to a variety of permutations. For brevity, I only report a few

such checks.12 For example, Panel A also reports estimates for a more parsimonious model of

PMAFE, with only DAGE andDBSIZEas control variables; the coefficients on bothDLOCAL

andDLOG are still significant. Univariate regressions and a variety of other untabulated speci-

fications produce identical inferences. Panel B reports statistics for monthly regressions, which

also produce similar results (t=-2.31 forDLOCAL, t=2.33 forDLOGD) in the baselinePMAFE

specification. Finally, when I run the accuracy tests on two-year ahead earnings per share fore-

casts instead of one-year forecasts, the coefficient estimate onDLOCAL is again negative and

significant, and the coefficient estimate onDLOGD is again positive and significant.

Since my results are similar for the two distance variables, I only report results for (and

limit my discussion to) theLOCALvariable for the remainder of the paper. My conclusions are

unchanged if I use theLOGD variable instead, or if I make small adjustments to the distance

threshold used in calculatingLOCAL.

Other results in Table 2 are worth noting. For example, the coefficient onAFFIL is negative

and significant in the accuracy regressions, indicating that affiliated analysts are more accurate

in their one-year earnings forecasts than other analysts. The magnitude of this effect is larger

than the proximity effect discussed above. However, untabulated statistics indicate that this result

is sensitive to the wayAFFIL is defined. For example, if I change the year cutoff, or define

affiliation only by IPO affiliation, this result is no longer significant. And yet, the coefficient

is always negative, calling into question the typical assumption that affiliated analysts produce

inferior forecasts.12Untabulated statistics indicate that local analysts also issue forecasts more frequently than non-local analysts,

and issue a higher percentage of lead forecasts than non-locals; these frequency and herding results are available onrequest.

12

Not surprisingly, the estimated coefficient onDPCTOPis always strongly positive, indicating

that analysts who are systematically more optimistic than other analysts tend to be less accurate.

I also find that analysts who work at high-status (HIGH) firms are more accurate than other

analysts; note, however, that this coefficient is only significant in thePMAFE specification. All

of the other control variables enter in the expected directions. For example, the coefficient on the

age of the forecast is positive and strongly significant in all specifications. In addition, consistent

with evidence presented in Clement (1999), analysts with more firm-specific experience and more

available resources are more accurate than other analysts (i.e., the coefficients onDFEXP and

DBSIZE are strongly negative in all specifications). As in Stickel (1992), I find that All-Star

analysts are more accurate than other analysts, although this finding is fairly weak in theDAFE

regressions.

ii. Bias Versus Informativeness

One difficulty with interpreting superior local accuracy as indicative of superior local information

is that the afore-mentioned accuracy tests do not distinguish bias from informativeness. For

example, local analysts may be more accurate simply because they are less optimistic, rather

than better informed.

I investigate this possibility by re-running the base-line yearly regressions depicted in Table

II after replacing thePCTOPvariable with an explicit control for the analyst’s relative optimism

about the stock in question, rather than his overall tendency towards optimism. Specifically, in the

DAFE specification,OPT is equal to the forecast error for analysti’s forecast of firmj for fiscal

yeart (FEi, j,t) minus the mean forecast error for firmj for fiscal yeart (FE j,t), where forecast

error is equal to the analyst’s latest forecast, minus actual company earnings (drawn from the

I/B/E/S Actuals File), as a percentage of stock price 12 months prior to the beginning of the fiscal

year.POPT is simplyOPT scaled by the absolute value of (FE j,t). Both of these variables thus

capture the direction of the forecast error.

Panel A of Table III shows that even after explicitly controlling for the relative optimism of

an analyst about a given stock, local analysts are still more accurate than other analysts. Specifi-

13

cally, the coefficient onDLOCAL is negative in both regressions, and significant (t=-2.17) in the

PMAFE specification. Further, when I replace the dependent variablesPMAFE andDAFE with

POPT andOPT, the coefficient onDLOCAL is no longer significant in either regression.

Panel B reports additional evidence on this issue. Specifically, I run monthly Fama and Mac-

Beth (1973) cross-sectional (and pooled) regressions of actual earnings on forecasted earnings,

for local analysts and non-locals separately. Panel B reveals strongly negative estimated inter-

cepts in these regressions. This is the classic analyst optimism bias reported frequently in the

literature (see, for example, DeBondt and Thaler (1990) and Abarbanell and Bernard (1992)).

Importantly, however, the estimated intercepts for the two samples (locals and non-locals) are

not significantly different; locals are not significantly less (or more) optimistic than non-locals.

On the other hand, the coefficient estimates (β) are significantly different for the two groups, as

are the R-squared values. In particular, the (β) estimate is closer to one and the R-squared value

is larger for local forecasts. Since the coefficient and R-squared in this type of regression tell

us about the informativeness of the forecast, these results indicate that local forecasts are more

informative than non-local forecasts, but not necessarily less optimistic.

iii. Sorting By Firm Characteristics

In this section I try to isolate areas and firm types for which the effect of geographic proximity

is particularly important. Table IV reports the estimated coefficients for theDLOCAL variable

only, for subsets of the Table II yearly regressions ofDAFE on the eight analyst characteristics.

I break down my original sample of 96,538 observations into a variety of subsamples. Each

column presents a subsample broken down by firm characteristics. For example, my first sort is

by size; each month I categorize stocks based on their one-month lagged market capitalization.

“SMALL” refers to stocks below the median level of market capitalization, and “LARGE” refers

to stocks above the median market capitalization.13 I also sort stocks each month based on their

lagged book-to-market ratio. Book equity (BE) is defined as the COMPUSTAT book value of

stockholders’ equity, plus balance sheet deferred taxes and investment tax credit (if available),

13Here and elsewhere, using terciles or quartiles to form these group designations, instead of medians, changes noconclusions.

14

minus the book value of preferred stock. Depending on availability, I use redemption, liquidation,

or par value (in that order) to estimate the book value of preferred stock. I match the yearly book

equity figure for all fiscal years ending in calendar yeart−1 with returns starting in July of year

t. This figure is then divided by market capitalization (ME) at montht−1 to form theBE/ME

ratio, so that theBE/ME ratio is updated each month. “GROWTH” stocks are those stocks below

the medianBE/ME, while “VALUE” stocks are those above the medianBE/ME. Finally, I sort

stocks into three groups based on past returns fromt−12 tot−2 (as in Fama and French (1996)).

I classify “LOSERS” as stocks below the median past return, and “WINNERS” as stocks above

the median past return.

Clear patterns emerge from these sorts. As seen in the first row of Table IV, the effect of

geographic proximity is concentrated in small stocks, value stocks, and stocks with poor past

returns. Local analysts covering these types of stocks are significantly more accurate than other

analysts. Further sorting firms into “Metro Stocks,” “Non-Metro Stocks,” and “Remote Stocks,”

as described in Section I, Table IV shows that local analysts are more accurate when covering

stocks located in remote cities or smaller cities. Again, these effects are concentrated in small

stocks and value stocks.14 The magnitude of the accuracy advantage is significantly larger for

local analysts covering stocks in remote areas ($0.054=-0.0629*0.860), particularly small stocks

in remote areas ($0.141=-0.0803*1.761).

Breaking the sample into two subperiods reveals that the effect of geographic proximity is

stronger in recent years. While the Internet and improvements in information technology may

have brought people closer together, they have not wiped out the advantage that local analysts

gain by being closer to the firms they cover. Interestingly, in the post-Reg FD sample the estimate

for DLOCAL is no longer significant, but the direction and magnitude is similar to the overall

estimate; thus the enactment of this law has not dramatically decreased the accuracy advantage

that local analysts exhibit. Finally, the magnitude of theDLOCAL coefficient is similar when

comparing a sample of forecasts issued within five days of a quarterly earnings announcement

date to a sample of all other forecasts, although only the non-earnings-related sample result is

significant.

14Untabulated industry breakdowns reveal little connection between the geographic concentration of an industryand the local accuracy advantage; these results are available on request.

15

iv. Sorting by Analyst Characteristics

Table V provides another breakdown of the baselineDLOCAL result from Table II, this time

focusing on analyst characteristics as opposed to firm characteristics. Each column in Table V

represents a different analyst type, such as non-New York City analysts (XNYC), All-Star analysts

(STAR), non-All-Stars (NON), analysts working at high-status firms (HIGH), and analysts not

working at high-status firms (LOW). Meanwhile, the rows in Panel A group observations by

the nature of the forecast, and the rows in Panel B classify analysts by the composition of their

portfolio.

Following Gleason and Lee (2002), Panel A categorizes forecast revisions according to their

signal attributes (SIGNAL). If a revision is above the analyst’s prior forecastandabove the prior

consensus forecast, I classify this forecast as signalling unambiguously good news (SIGNAL=1).

By contrast, if a revision is below the analyst’s prior forecastand below the prior consensus

forecast, I classify this forecast as signalling unambiguously bad news (SIGNAL=-1). I place all

other forecasts into a third category (SIGNAL=0).

Panel A shows that even after removing all New York analysts, the local accuracy advantage

persists (DLOCAL=-0.0250,t=-2.76).15 Meanwhile, the magnitudes of theDLOCALcoefficient

suggest that the local effect is strongest among All-Star analysts and high-status analysts, al-

though the coefficient is only significant for low-status analysts (t=-2.07). Finally, local analysts

appear to be considerably more accurate when making unambiguous downward revisions. This

last result is strongly confirmed in the sub-samples of non-New York analysts, non-All-Stars, and

low-status analysts, and present (but insignificant) in the samples of All-Stars and high-status

analysts.

To form the coverage terciles depicted in Panel B, I simply compute the number of stocks that

each analyst covers in a given year: “Low-Coverage” analysts are those who fall below the 33rd

percentile of number of stocks covered by an analyst in a given year, while “High-Coverage”

analysts are those above the 66th percentile. I further categorize analysts by the share of local

stocks in their portfolio; specifically,LOCMED is a dummy variable equal to one if the share of

15Similarly, after dropping all analysts located within 100 kilometers of San Francisco, the coefficient onDLOCALis comparable (-0.0230) and still significant (t=-2.17) in thePMAFE yearly specification from Table II.

16

local stocks in an analyst’s portfolio is above the median share of local stocks over all analysts’

portfolios. Panel B shows that high coverage analysts exhibit the strongest local advantage, par-

ticularly those whose portfolios are tilted towards local stocks (DLOCAL=-0.0512,t=-2.67). This

result suggests that the local accuracy advantage may have more to do with private information

(e.g., absorbing the local culture or running in the same social circles as local business leaders),

and less to do with effort (i.e., the reduced cost of information gathering for local analysts).

v. Endogenous Coverage Decisions

Looking at coverage in the preceding way masks the fact that analysts are not randomly assigned

to firms and how intensively to follow them. Analysts make their coverage decisions endoge-

nously, and these coverage decisions could lead naturally to a link between performance and

proximity in the absence of asymmetric information costs. I explore three versions of this hy-

pothesis in this section.

One possibility is that local analysts, particularly those in remote areas, simply cover fewer

firms, and hence benefit from the increased concentration of their attention relative to metropoli-

tan analysts. Panel A of Table VI shows that local analysts and remote analysts do indeed cover

fewer firms on average (11.87 and 11.10 firms per year, respectively, compared to 12.37 for the

entire population of analysts). However, Panel B of Table VI shows that when I include a variable

equal to the number of firms covered by each analyst (DTASK) in the original regression spec-

ification from Table II, local analysts still exhibit an accuracy advantage (t=-1.92 for the entire

sample, andt=-2.51 for remote analysts). Further, as an information story would suggest, remote

analysts actually cover more local firms (2.22 per year) than the average analyst (2.01). Due to

the fixed costs of relocating and establishing local ties, it is likely to be too costly for analysts in

large cities to simply relocate to remote areas.

Another possibility is that local analysts may be more specialized (e.g., by industry) than

other analysts. Specialists might be more accurate than broader analysts even if the cost of

information is the same across locations, producing a relation between distance and accuracy. I

explore this issue by examining the performance of “Expert” analysts, which I define as analysts

17

covering firms in only one 2-digit SIC code in a given year.16 Panel A of Table VI shows that

expert analysts are actually located further from the firms they cover on average and tend to cover

fewer local stocks (1.67 per year) than the average analyst (2.01). Both remote analysts and local

analysts tend to be less specialized (31.63% and 32.51% are experts, respectively, compared to

32.66% for the entire sample), suggesting that specialization is not a major determinant of local

analyst performance. As shown in Panel B of Table VI, including a dummy variable equal to

one if the analyst is a specialist (DSPEC) in the regression from Table II does not diminish the

predictive power of proximity.

A final version of the endogenous coverage hypothesis builds on the first story, but argues

that what may drive the local analyst advantage is not the total number of firms covered, but

rather the intensity with which certain individual firms are covered. Since fund managers and in-

dividual investors are biased towards holding local stocks (see Coval and Moskowitz (1999) and

Zhu (2002)), the customers of local analysts may be driving the proximity patterns documented

above. Local analysts may simply spend more time analyzing local stocks, since these are the

stocks that their local clients demand. Testing this idea requires a measure that is correlated with

local investor demand but which does not necessarily provide information to analysts. One such

measure is advertising intensity, since Zhu (2002) shows that individual investors tend to invest

more in stocks that advertise heavily, and individuals may be more likely to gain exposure to the

advertising of local companies (Nelson (1974). Breaking the sample into firms with high and

low advertising intensity (defined as advertising expense divided by sales), Panel B of Table VI

shows that local analysts do perform much better when covering high advertising firms than low

advertising firms (DLOCAL=-0.0510 witht=-3.25 compared toDLOCAL=0.0377,t=1.63)), al-

though this result is much weaker for remote firms. Conditioning on advertising intensity greatly

reduces the sample size, however, as roughly eighty percent of the forecasts in the sample come

from firms with zero advertising expense. Among firms with zero advertising expense, the local

accuracy advantage persists. Nevertheless, I cannot completely rule out the possibility that visi-

bility may play a role in the performance of local analysts; a full exploration of this issue would

16These results are robust to a variety of definitions of analyst specialization, such as the number of 2-digit SICcodes covered or the dispersion in 2-digit SIC codes within an analyst’s portfolio, and to different industry classifi-cations (e.g., 4-digit SIC codes or the 17-48 industry portfolios of Fama and French (1997)).

18

require merging individual investor holdings data with analyst forecast data, a task beyond the

scope of this paper.

III. Stock Price Impact of Forecast Revisions

A. Test Design

Another way to test if local analysts bring new and valuable information to the market is by an-

alyzing the stock price impact of forecast revisions. Since weak analysts can simply mimic the

earnings forecasts of timely skilled analysts in order to improve the accuracy of their forecasts,

relative forecast accuracy is not a sufficient statistic for testing if geographic proximity facili-

tates information flow.17 Specifically, I test the hypothesis that forecast revisions made by local

analysts have a greater impact on stock prices than revisions made by distant analysts.18 Such

a finding would provide strong evidence that local analysts have an information advantage over

other analysts. This type of test also helps gauge theeconomicsignificance of my prior results.

I test the hypothesis that forecast revisions made by local analysts have a greater impact

on stock prices by running Fama and MacBeth (1973) cross-sectional regressions of three-day

average excess returns (AARi,−1,1) on the magnitude of the revision (see Beaver, Clarke, and

Wright (1979)), firm size (see Atiase (1985)) and dummy variables indicating affiliation status

and local status. The following regression is used:

AARj,−1,1 = β0 +β1LNMEj,−30+β2SUFi, j,0 +β3AFFILi +β4LOCALi + ε (2)

whereAARj,−1,1 is the average size-adjusted announcement return for firmj, LNMEj,−30 is

the (log of) market capitalization of the stock one month prior to the revision,SUFi, j,0 is the

standardized unexpected forecast made by analysti for firm j, AFFILi is a dummy variable

17Cooper, Day, and Lewis (2001) provide evidence that ranking analysts solely on forecast accuracy can lead tomisclassification errors.

18While Carleton, Chen, and Steiner (2002) conduct a similar test of the market’s response to regional versusnational analysts’ announcements, they do not focus on geographic proximity as their unit of analysis, but rather theperceived reputation of the analyst’s employer.

19

equal to one if the analyst is an affiliated analyst as defined in Table II (and zero otherwise), and

LOCALi is a dummy variable equal to one if the analyst is a local analyst (and zero otherwise).

The standardized unexpected forecast equals analysti’s forecast for firmj on day 0 minus analyst

i’s prior forecast for firmj (in fiscal yeart), scaled by the cross-sectional standard deviation of

all prior outstanding forecasts for firmj.19

I break the regressions down into two types of revisions, one group signalling “good” news

and the other signalling “bad” news. As in Panel A of Table V, I categorize revisions according

to their signal attributes (SIGNAL). However, for these tests I focus only on those forecasts

signalling unambiguously good news (SIGNAL=1) or unambiguously bad news (SIGNAL=-1),

and exclude all forecasts with conflicting signals (SIGNAL=0).20 I also restrict these regressions

to various subsamples according to the location of the firm being covered. As before, I group

stocks located in large cities and small cities separately to see if local analysts have a greater

impact on stock prices for particular types of firms. I also divide the sample up into different

analyst types (as in Table V) and different subperiods (as in Table IV).

An advantage of equation (2) is that the estimated coefficient for the dummy variable rep-

resents the marginal abnormal return associated with unambiguous forecast revisions by local

analysts. I run these regressions monthly, with revisions placed in the calendar month in which

day 0 falls, and compute the time-series average of the 85 monthly estimated coefficients.

I also analyze the relation between forecast revisions and excess stock returns during the post-

release period. For this test I substitute the post-release average excess return from day 2 to day

64 (AARj,2,64) in place of the three-day average excess return in equation (2). This test provides

direct evidence concerning the speed with which investors react to the information contained in

analysts’ forecast revisions.

19As in Stickel (1992), when computingSUFi, j,0 I set a divisor less than $.25 arbitrarily equal to $.25 to mitigatesmall denominators, and scaled forecast revisions are truncated at -200% and +200%. Scaling forecast revisions bythe absolute value of the prior forecast or by price changes no conclusions.

20I obtain similar results if I restrict the regressions to revisions of a relatively constant magnitude ofSUF (e.g.,the top 10%SUF or the bottom 10%SUF) as in Stickel (1992), rather than according to signal attributes.

20

B. Results

Table VI reports the time-series average of the estimated coefficients for Fama and MacBeth

(1973) cross-sectional regressions of equations (2). Panel A reports the regression results for

equation (2) and shows that both strong positiveand strong negative revisions by local analysts

affect stock prices more than other analysts. For the entire sample of stocks, covering 86,127

revisions, local revisions are associated with an incremental average excess return of 0.095%

per day in the three days surrounding strong positive revisions, which is significant at the five

percent level. This effect is slightly larger for stocks located in small cities (0.102% per day).

For large negative revisions, the effects are even stronger (-0.128% andt=-2.46 for all stocks,

-0.164% andt=-2.54 for non-metro stocks). The fact that the market’s response to local revisions

is stronger on the downside is consistent with the finding in Table V that the local accuracy

advantage is strongest for large downward revisions. Overall, the finding that both positive and

negative revisions by local analysts are met with significant price responses provides support for

the notion that geographically proximate analysts are better informed than other analysts.

Panel C shows that the price responses associated with local revisions more than doubled in

magnitude from the early period to the more recent period (0.057% up to 0.126% for positive

revisions, and -0.058% to -0.180% for negative revisions). Again this finding is consistent with

the result reported in Table IV that the local accuracy advantage has increased in recent years.

The post-Reg FD results are even more dramatic, as negative revisions by locals in the 13 months

after the enactment of the law were met with an incremental average excess return of -0.421%

per day in the three-day announcement period.

The results in Panel D indicate that strong positive revisions by local non-New York analysts,

local All-Star analysts, and local high-status analysts are met with significant incremental price

responses, but strong negative revisions are not. This is the first piece of evidence that conflicts

with the results reported in Table V, since the local accuracy advantage for all three of these ana-

lyst types was found to be strongest for downward revisions; note, however, that these particular

accuracy results were insignificant for the All-Star and high-status samples. The large negative

price response associated with local low-status analysts is again consistent with the finding in

21

Table V that these particular analysts are more accurate on the downside. Finally, the incremen-

tal price responses for locals covering high advertising firms are insignificant (albeit for a much

smaller sample), calling into question the robustness of the earlier finding that the local accuracy

advantage is particularly strong among firms of this type.

The affiliation results in Table VI are striking. Panel A indicates that the market’s response to

unambiguously negative revisions by affiliated analysts is quite large. Consistent with an agency

story, affiliated analysts are associated with an incremental average excess return of -0.555%

per day (t=-3.67) in the three days surrounding strong negative revisions, but only 0.074% per

day (t=0.68) around strong positive revisions. Since affiliated analysts may have an incentive

to release positive news and a disincentive to issue negative forecasts, the market views positive

revisions by affiliated analysts as uninformative and negative revisions as extremely informative.

Not surprisingly, these affiliation patterns are even stronger in recent years. For example, as

shown in Panel C, strong positive revisions by affiliated analysts in the more recent subperiod

were met with an incremental average excess return of -0.190% per day (albeit insignificant),

indicating that the market was completely discounting good news by affiliated analysts during

this period. However, the market still viewed strong negative revisions by affiliated analysts as

informative during this period, as these were met with an incremental average excess return of

-0.665% per day.

Panel D reports that this asymmetry in the market’s response to affiliated analysts’ revisions

is not found among non-New York analysts, non-All-Stars, and low-status analysts. Since these

are the three groups where one might expect agency problems to be less severe, this result is not

surprising.

The regression results in Panel B indicate that the price responses documented above are

short-term in nature. Although the signs are the same as in Panel A, there is no significant

differential relation between forecast revisions and post-release excess returns for local analysts

versus non-local analysts, or for affiliated analysts versus non-affiliated analysts.

In summary, both positive and negative revisions by local analysts have a greater stock price

impact than revisions by non-locals, and this difference is strongest for non-metro stocks, and

22

in the more recent subperiod. This additional impact continues on for the next three months,

but the long-run differential is not significant. By contrast, only negative revisions by affiliated

analysts have a greater stock price impact than revisions by unaffiliated analysts. Consistent with

an agency story, this asymmetry is strongest in metro stocks, All-Star analysts, and high-status

analysts.

IV. Analysts’ Stock Recommendations

In this section I analyze the effect of distance on analysts’ stock recommendations. Rather than

simply replicating the same (or similar) tests as those in Sections II-III on the recommendation

data, I use geography to shed light on a specific phenomenon that has received considerable at-

tention in recent years with respect to stock recommendations, namely the underwriter affiliation

bias. Much of the empirical evidence documenting an affiliation bias in analysts’ research has fo-

cused solely on stock recommendations, not earnings forecasts. Indeed my accuracy tests found

no evidence of inferior performance by affiliated analysts in their one-year earnings forecasts.21

A. Stock Price Impact of Analyst Recommendations: Local Versus Affiliated An-

alysts

My first test replicates a commonly-cited example of the affiliation bias in stock recommenda-

tions, and re-examines this result by conditioning on the location of the analyst. As in Lin and

McNichols (1998), I examine the market’s reaction to affiliated and unaffiliated analysts’ stock

recommendations.

I analyze the stock price impact of affiliated analysts versus unaffiliated analysts by running

Fama and MacBeth (1973) cross-sectional regressions each month of three-day average excess

21Similarly, Teoh and Wong (2002) find no evidence that the predictive power of accruals to explain forecastserrors differs by affiliation status.

23

returns on a series of dummy variables. I estimate the relation between returns and recommenda-

tions as follows:

AARj,−1,1 = β1SBj,A +β2B j,A +β3H j,A +β4SBj +β5B j +β6H j +β7Sj + ε (3)

whereAARj,−1,1 is the average size-adjusted announcement return for firmj, SB is a dummy

variable equal to one if the recommendation is a “Strong Buy” and zero otherwise,B is dummy

variable equal to one if the recommendation is a “Buy” and zero otherwise,H is a dummy variable

equal to one if the recommendation is a “Hold” and zero otherwise, andS is a dummy variable

equal to one if the recommendation is a “Sell” or ”Strong Sell” and zero otherwise. The variables

subscripted byA equal one if the respective recommendation is issued by an affiliated analyst, and

zero otherwise. I define an affiliated analyst as an analyst who issues a recommendation on a stock

in the two years after the issuance of an secondary equity offering (SEO) of that stock, where the

lead underwriter for that SEO is also the brokerage firm that employs the analyst. As in Lin

and McNichols (1998), I estimate the model for the sample of all recommendations issued in the

two years after the SEO. Because there are very few affiliated “Sell” recommendations, these are

grouped together with affiliated “Hold” recommendations. Dropping this restriction changes no

conclusions. The coefficientsβ1, β2, andβ3 thus capture the mean incremental returns associated

with affiliated analysts’ recommendations relative to the same recommendations by unaffiliated

analysts, and the t-statistics for these coefficients indicate the significance of the difference in the

mean market reactions for affiliated and unaffiliated analysts.

Panel A of Table VII presents my results, which are very similar to those originally reported

in Lin and McNichols (1998). The estimation results in the first row of Panel A indicate a signifi-

cant positive response to unaffiliated “Strong Buy” recommendations (t=16.11), and a significant

negative response to unaffiliated “Hold” (t=-14.23) and “Sell” recommendations (t=-7.37). This

evidence indicates that investors view analysts’ recommendations as informative. Particularly, in-

vestors seem to undo the bias in unaffiliated “Hold” recommendations, since they interpret these

recommendations as negative rather neutral information about a stock.

24

The coefficient onβ1 is not significant, suggesting that investors do not view affiliated “Strong

Buy” recommendations as less informative than unaffiliated “Strong Buy” recommendations.

By contrast, the coefficient onβ2 is negative and significant, although the significance of this

particular result is not robust to the cutoff period used (here I include all recommendations issued

in the two years after the SEO). The coefficientβ3, however, is strongly significant is all of my

specifications (t=3.90), and is large in magnitude (-1.76% per day). This is the central result in

Lin and McNichols (1998), and it appears in my sample as well, namely that the market interprets

an affiliated “Hold” to mean “Sell” to a greater degree than an unaffiliated “Hold.” This result

suggests that affiliated analysts are more likely to issue a “Hold” recommendation when a “Sell”

recommendation is warranted.

I then explore the effect of geographic proximity in this context. Panel B of Table VII reports

that affiliated analysts that are also local arenot associated with a significant negative incremen-

tal price response (-0.25% per day,t=-0.78) when they issue “Hold” recommendations. Thus

the market does not view an affiliated local “Hold” recommendation as more informative than an

unaffiliated “Hold” recommendation. However, the reaction to “Strong Buy” recommendations

is now positive and significant (0.32% per day,t=1.96), suggesting that local affiliated analysts

do have an information advantage over other analysts when issuing these particular recommen-

dations.

Panel C shows that distant affiliated “Hold” recommendations are met with an incremental

price reaction of -1.74% per day on average over the three-day announcement period. Thus the

agency cost estimates implied by Panel A are only borne out for distant affiliated analysts. A

possible explanation for this finding is explored in the next section.

I also examine the relation between pre-release excess returns and stock recommendations

by substituting the pre-release average excess return from day−22 to day−2 (AARj,−22,−2) in

place of the three-day average excess return in equation (5). The second row of Panel A reports a

result similar in spirit to the “booster shot” result first reported in Michaely and Womack (1999).

Using a sample of IPOs in 1990-1991, Michaely and Womack (1999) report that returns of firms

with affiliated “Buy” recommendations declined, on average, 1.6% in the 30 days prior to the

“Buy” recommendation, while firms receiving unaffiliated “Buy” recommendations increased

25

4.1% over the same period. They conclude that affiliated analysts attempt to boost the stock

prices of firms they have taken public at times when such a booster shot is really need, i.e., when

the firm is performing poorly. Similarly, the second row of Panel A shows that in my sample

of SEOs, affiliated “Buy” recommendations are associated with an incremental average excess

return of -0.11% per day (t=-1.88) in the pre-release period relative to “Buy” recommendations

by unaffiliated analysts.

As with the “Hold” affiliated announcement effect, this booster shot effect is concentrated

among distant affiliated analysts. The second row of Panel B reports an incremental average

excess return of only 0.01% per day (t=0.07) in the pre-release period for local affiliated “Buy”

recommendations relative to “Buy” recommendations by unaffiliated analysts. By contrast, Panel

C shows that distant affiliated “Buy” announcements are met with an incremental average excess

return of -0.09% per day (t=-1.75) in the pre-release period.

B. Optimism in Analyst Recommendations: Local Versus Affiliated Analysts

The evidence that local affiliates appear to issue more informative recommendations than distant

affiliates in the “Strong Buy” category is consistent with an information advantage on the part of

local affiliates. However, the finding that local affiliates’ “Hold” recommendations appear unbi-

ased relative to their distant counterparts is puzzling, and motivates a more extensive analysis.

I examine the robustness of this last result by employing the matching procedure described

in Lin and McNichols (1998) to compare the stock recommendations of affiliated analysts and

unaffiliated analysts. I define an affiliated analyst as an analyst who issues a recommendation on

a stock in a designated time period around the issuance of an secondary equity offering (SEO) of

that stock, where the lead underwriter for that SEO is also the brokerage firm that employs the

analyst.22 I then find a recommendation for this same stock by an unaffiliated analyst within sixty

days of the affiliated recommendation; only recommendations with an eligible match are included

in the test. When there is more than one affiliated recommendation for an underwriter-offering

observation, the recommendation made on the day closest to the offering date is included in the

22Table VIII presents results for a variety of different designated time periods.

26

sample. Similarly, when there is more than one unaffiliated recommendation within sixty days of

the respective affiliated recommendation, I choose the unaffiliated recommendation issued most

closely to the date of the affiliated recommendation. This research design therefore controls for

differences in the characteristics of firms that affiliated versus unaffiliated firms choose to cover.

Panel A of Table VIII presents the mean differences between pairs of affiliated and unaffiliated

recommendations. Using 247 matched pairs of recommendations, I find that the mean affiliated

recommendation is 1.725 (1=“Strong Buy,” “5=Strong Sell”), while the mean unaffiliated rec-

ommendation is 1.895; this difference (-0.170) is statistically significant at the five percent level,

indicating that affiliated analysts are significantly more optimistic than unaffiliated analysts for

this sample of recommendations. This result is very similar to the findings in Lin and McNichols

(1998), although they analyze a different time period (SEOs from 1989-1994). Again this result is

typically interpreted as signalling an agency problem on the part of equity analysts; the argument

is that analysts bias their recommendations upwards for stocks underwritten by their brokerage

house, since analysts’ pay is sometimes tied to investment banking revenues. The second column

of Panel A shows that this affiliation bias is concentrated in stocks located in large cities: the

difference between affiliated recommendations and unaffiliated recommendations is -0.390 for

metro stocks, which is again strongly significant. By contrast, the sample of non-metro stocks

reveals a much smaller bias (-0.120), which is no longer statistically significant. When I restrict

the sample to include only affiliated analysts who produce recommendations within 90 days of

the SEO, the mean difference is larger and still significant (-0.231,t=-2.81), but the number of

matches is smaller (147).

After verifying that the affiliation bias documented in Lin and McNichols (1998) is again

present in my sample, I examine the effect of geographic proximity in this setting. Specifically,

I look to see if local affiliates’ recommendations are unbiased, as my previous stock price tests

suggest. To do so, I employ the same matching procedure described above, except that I now

group local and non-local affiliated analysts separately. Panel A shows that affiliated analysts

that arealso local tend to issue recommendations that arenot significantly more optimistic than

a matched sample of unaffiliated analysts. The strongly significant affiliation bias documented in

Panel A is not present for affiliated analysts who are also local. A similar result appears in the

27

sample of metro stocks, although this finding is based on only 6 matched pairs. By contrast, the

last few rows of Panel A show that the affiliation bias documented above is concentrated among

distant affiliates (mean difference=-0.181,t=-2.41).

This result, coupled with the similar findings on price impact in the prior section, raises the

question of why local affiliated recommendations are unbiased. Panel B provides some summary

statistics to help explore this question. A few obvious differences between the sample of local

affiliates and the sample of distant affiliates stand out, the first being the differing percentage of

high-status analysts between the two groups (34.21% for local affiliates and 55.35% for distant

affiliates). The large percentage differences in All-Star analysts (15.79% to 30.23%) and New-

York analysts (31.58% to 68.84%) between the two groups paint a similar picture. Clearly local

affiliated analysts are fundamentally different than non-local affiliated analysts.

Table IX presents some summary statistics for the entire sample of recommendations to see

if these patterns are widespread. Indeed, the percentage of high-status recommendations among

distant affiliated recommendations (49.43%) is far higher than among local affiliated recommen-

dations (30.56%). Further, the percentage of high-status recommendations among affiliated rec-

ommendations as a whole is far higher than the percentage in the entire population of recommen-

dations (46.46% to 21.09%). These findings suggest that local affiliates may be unbiased simply

because they tend to represent a different class of analyst, one that is less exposed to the agency

problem often cited to explain the affiliation bias. Since local affiliates account for only 14.90%

of all affiliated recommendations, and account for a higher percentage of first-time offerings than

repeat offerings (15.81% to 14.36%), local affiliates may have less incentive to bias their recom-

mendations upwards in the hopes of garnering additional investment banking business.

Panel B of Table IX explores this idea further. As before, I examine differences in mean

recommendation codes. However, in this case the tests are performed on various subsets of the

analyst population. The variableX is associated with the analyst characteristic in the desig-

nated column. For example,AFFILX=1 refers to affiliated recommendations where the affiliated

analyst is also in categoryX; the first row of theHIGH column thus reports the mean recommen-

dation code for affiliated analysts who also work at high-status firms. All affiliated recommen-

dations are matched to non-affiliated recommendations as described in Table VIII. The first few

28

rows of Panel B show that the affiliation bias is concentrated among high-status analysts (mean

difference=-0.252,t=2.64) and analysts located in New York city (mean difference=-0.194,t=-

2.29); since 88.6% of high-status analysts are located in New York, these similarities are not

surprising.

The last few rows of Panel B report an interesting result. In these rows I form matched

pairs strictly by analyst characteristics. For example, instead of matching affiliated analysts with

unaffiliated analysts, I simply match up high-status analysts (whether or not they are affiliated

with the SEO) with low-status analysts using the same timing rules around SEO issuances as

before. Using this approach I find a mean recommendation difference equal to -0.083 (t=-1.24)

in the high-status minus low-status match. High-status analysts thus appear more optimistic than

low-status analysts around SEO issuances, even when they are not explicitly affiliated with the

deal. Further, when I limit the sample to high-status analysts producing recommendations within

90 days of the SEO date, this mean recommendation difference climbs to -0.201 (t=-2.48), a

figure almost as large as the original affiliation bias figure reported in Panel A of Table VIII (=-

0.231,t=-2.78)! This finding suggests that the affiliation bias in stock recommendations is largely

a high-status analyst phenomenon. Since high-status affiliates are more likely to be located in

New York and further away from the stocks they cover, the results in Tables VII and VIII that

local affiliates are unbiased is not surprising.

V. Conclusion

This paper is the first to connect the concept of geography to U.S. equity analysts’ forecasts and

recommendations. The effect of distance has already been documented in corporate governance,

delegated portfolio management, and many other areas of economics, but has not been adequately

examined in the context of equity analysis. I find evidence in support of the hypothesis that

geographically proximate analysts possess an information advantage over other analysts. Local

analysts are more accurate and have a greater impact on security prices than other analysts. These

effects are strongest in recent years, for firms located in small cities and remote areas, and for

analysts issuing downward revisions. I find little support for the notion that the link between

29

geography and performance is a natural by-product of the endogenous coverage decisions made

by analysts, although I do find that local investor demand or visibility may play a role in local

abnormal performance. Analyzing the flow of information between investors and analysts by

merging data on fund manager holdings, individual investor holdings, and analyst forecast data

would provide an ideal testing ground to evaluate this last possibility, and is a subject for future

research.

I find no evidence of an underwriter affiliation bias for affiliated analysts that are also lo-

cated nearby the firm being covered. I argue that this is because local affiliates are less likely

to be working at high-status firms, where agency problems are intense due to the constant pres-

sure to garner investment banking business. This suggests that their is an important geographic

component to the debate about agency problems in the analyst industry.

Since conflicts of interest appear to be particularly problematic when affiliated analysts are

located far away from the firm, an interesting extension of this paper would be to analyze the

effect of switching locations on analyst (particularly high-status analyst) performance. Broader

questions such as which factors drive location changes for analysts, and the extent to which

knowledge is portable are also intriguing. These and other issues are left to future research.

30

References

Abarbanell, Jeffrey S., and Victor L. Bernard, 1992, Tests of analysts’ overreaction/underreaction

to earnings information as an explanation for anomalous stock price behavior,The Journal of

Finance47, 1181–1207.

Agrawal, Anup, and Sahiba Chadha, 2002, Who is afraid of Reg FD? The behavior and perfor-

mance of sell-side analysts following the SEC’s Fair Disclosure rules, Working paper, Univer-

sity of Alabama.

Atiase, Rowland K., 1985, Predisclosure information, firm capitalization, and security price be-

havior around earnings announcements,Journal of Accounting Research23, 21–36.

Bacmann, Jean-Francois, and Guido Bolliger, 2001, Who are the best? Local versus foreign

analysts on the Latin American stock markets, Working paper, University of Neuchatel.

Bailey, Warren, Haitao Li, Connie X. Mao, and Rui Zhong, 2003, Regulation fair disclosure

and earnings information: Market, analyst, and corporate responses,Journal o f Finance,

forthcoming.

Beaver, William, Roger Clarke, and William Wright, 1979, The association between unsystem-

atic security returns and the magnitude of the earnings forecast error,Journal of Accounting

Research17, 316–340.

Berger, Allen N., Nathan H. Miller, Mitchell A. Petersen, Raghuram G. Rajan, and Jeremy C.

Stein, 2002, Does function follow organizational form? Evidence from the lending practices

of large and small banks, Harvard University working paper.

Bolliger, Guido, 2001, The characteristics of individual analysts’ forecasts in Europe, FAME

Research Paper Series, No. 33.

Carleton, Willard T., Carl R. Chen, and Thomas L. Steiner, 2002, Are all security analysts equal?,

Journal o f Financial Research, forthcoming.

Chang, Charles, 2002, Information footholds: expatriate analysts in an emerging market, Work-

ing paper, Haas School of Business, U. C. Berkeley.

31

Chevalier, Judith, and Glenn Ellison, 1999, Career concerns of mutual fund managers,Quarterly

Journal of Economics114, 389–432.

Clement, Michael B., 1998, Some considerations in measuring analysts’ forecasting perfor-

mance, Working paper, University of Texas.

Clement, Michael B., 1999, Analyst forecast accuracy: Do ability, resources, and portfolio com-

plexity matter?,Journal of Accounting and Economics27, 285–303.

Cooper, Rick A., Theodore E. Day, and Craig M. Lewis, 2001, Following the leader: A study of

individual analysts’ earnings forecasts,Journal of Financial Economics61, 383–416.

Coval, Joshua D., and Tobias J. Moskowitz, 1999, Home bias at home: Local equity preference

in domestic portfolios,The Journal of Finance54, 2045–2073.

Coval, Joshua D., and Tobias J. Moskowitz, 2001, The geography of investment: Informed trad-

ing and asset prices,Journal of Political Economy4, 811–841.

DeBondt, Werner F. M., and Richard H. Thaler, 1990, Do security analysts overreact?,American

Economic Review80, 52–57.

Dechow, Patricia M., Amy P. Hutton, and Richard G. Sloan, 1998, The relation between ana-

lysts’ forecasts of long-term earnings growth and stock price performance following equity

offerings, Working paper, University of Michigan.

Diether, Karl B., Christopher J. Malloy, and Anna Scherbina, 2002, Differences of opinion and

the cross-section of stock returns,The Journal of Finance57, 2113–2140.

Dugar, A., and S. Nathan, 1995, The effects of investment banking relationships on financial

analysts’ earnings forecasts and investment recommendations,Contemporary Accounting Re-

search12, 131–160.

Fama, Eugene F., and Kenneth R. French, 1996, Multifactor explanations of asset pricing anoma-

lies,The Journal of Finance51, 55–84.

Fama, Eugene F., and Kenneth R. French, 1997, Industry costs of equity,Journal of Financial

Economics43, 153–193.

32

Fama, Eugene F., and James MacBeth, 1973, Risk, return and equilibrium: empirical tests,Jour-

nal of Political Economy51, 55–84.

Gleason, Cristi A., and Charles M. C. Lee, 2002, Analyst forecast revisions and market price

formation, Working paper, University of Arizona.

Grinblatt, Mark, and Matti Keloharju, 2001, Distance, language, and culture bias: The role of

investor sophistication,The Journal of Finance56, 1053–1073.

Hong, Harrison, and Jeffrey D. Kubik, 2003, Analyzing the analysts: Career concerns and biased

earnings forecasts,The Journal of Finance58, 313–351.

Hong, Harrison, Jeffrey D. Kubik, and Jeremy C. Stein, 2002, Thy neighbor’s portfolio: word-

of-mouth effects in the holdings and trades of money managers, Working paper, Stanford

University.

Huberman, Gur, 2001, Familiarity breeds investment,Review of Financial Studies14, 659–680.

Lamont, Owen, 2002, Macroeconomic forecasts and microeconomic forecasters,Journal of Eco-

nomic Behavior and Organization48, 265–280.

Lin, Hsiou-wei, and Maureen F. McNichols, 1998, Underwriting relationships, analysts’ earnings

forecasts and investment recommendations,Journal of Accounting and Economics25, 101–

127.

Michaely, Roni, and Kent L. Womack, 1999, Conflict of interest and the credibility of underwriter

analyst recommendations,Review of Financial Studies12, 653–686.

Nelson, Phillip, 1974, Advertising as Information,Journal of Political Economy82, 729–754.

Orpurt, Steven, 2002, Analyst location and forecast accuracy: International evidence about expert

analysts and asymmetric information, Working paper, University of Chicago.

Payne, Jeff L., and Wayne B. Thomas, 2002, The implications of using stock-split adjusted

I/B/E/S data in empirical research, Working paper, University of Oklahoma.

33

Petersen, Mitchell A., and Raghuram G. Rajan, 2002, Does distance still matter? The information

revolution in small business lending,The Journal of Finance57, 2533–2570.

Portes, Richard, and Helene Rey, 2002, The determinants of cross-border equity transaction

flows, Working paper, London Business School.

Stickel, Scott E., 1992, Reputation and performance among security analysts,The Journal of

Finance47, 1811–1836.

Teoh, Siew Hong, and T. J. Wong, 2002, Why do new issuers and high-accrual firms underper-

form: The role of analysts’ credulity,Review o f Financial Studies, forthcoming.

Zhu, Ning, 2002, The local bias of individual investors, Working paper, Yale University.

34

Table ISummary Statistics for U.S. Equity Analysts

This table reports summary statistics for the two main samples used in this paper. Panel A reports statistics forthe one-year earnings forecast sample, and Panel B reports statistics for the stock recommendation sample. Bothsamples are broken down by forecast as well as firm characteristics. Forecasts are classified as “Metro” if the analystis located in one the 20 most populated cities, and “Remote” if the analyst is not located within 400 kilometersof any of the 20 most populated cities; firms are classified analogously but according to the location of the firm’sheadquarters rather than the location of the analyst. A forecast is classified as “Local” if the analyst is located within100 kilometers of the headquarters of the firm being covered. Analysts and firms located in Hawaii and Alaska areexcluded from the sample, and the District of Columbia is counted as a state. For each forecast, the brokerage size(“Brok. Size”) is equal to the number of analysts working for the same firm in the same calendar year as the analystproducing the forecast. Analyst coverage (“Coverage”) is equal to the number of analysts covering the firm eachfiscal year.

Panel A: One-Year Earnings Forecast Sample (n=348,571)Forecast Type No. of No. of Avg. Dist. Pct. Pct. Mean (Median)

Analysts States To Firms (km) Local NYC Brok. SizeAll 4254 34 1515.60 16.84 56.47 37.44 (35)Metro 3358 13 1602.68 17.07 68.67 42.58 (41)Non-Metro 1141 33 1179.84 15.95 0.00 17.61 (12)Remote 462 13 1054.66 22.25 0.00 14.88 (10)199411-199712 2645 33 1474.77 17.02 59.08 30.11 (28)199801-200112 3849 33 1545.60 16.71 54.55 42.83 (41)Firm Type No. of No. of Avg. Dist. Avg. Size Avg. No. of Mean (Median)

Firms States To Analysts (km) ($ mill.) Employees CoverageAll 4344 49 1394.69 1824.43 6947.3 5.79 (3.83)Metro 877 14 1505.53 2290.81 8327.7 7.00 (4.60)Non-Metro 3544 49 1373.95 1711.51 6650.0 5.49 (3.63)Remote 966 28 1415.26 1521.59 5927.2 5.26 (3.67)199411-199712 2998 49 1326.44 1329.70 7523.8 5.47 (3.75)199801-200112 3610 48 1419.47 2586.19 7902.2 6.01 (4)

Panel B: Stock Recommendation Sample (n=87,001)Recommendation Type No. of No. of Avg. Dist. Pct. Pct. Mean (Median)

Analysts States To Firms (km) Local NYC Brok. SizeAll 4054 34 1509.76 16.83 49.41 40.33 (27)Metro 3183 13 1605.68 17.36 64.00 44.96 (34)Non-Metro 1101 33 1227.30 15.27 0.00 26.70 (13)Remote 436 13 1168.93 21.85 0.00 23.22 (10)199411-199712 2204 33 1447.99 17.14 49.86 28.88 (22)199801-200112 3448 33 1545.67 16.65 49.14 46.99 (33)Firm Type No. of No. of Avg. Dist. Avg. Size Avg. No. of Mean (Median)

Firms States To Analysts (km) ($ mill.) Employees CoverageAll 4271 48 1403.62 2160.74 7632.0 3.26 (2.33)Metro 903 14 1557.33 3037.81 8875.5 3.64 (2.67)Non-Metro 3457 48 1372.69 1931.27 7338.8 3.15 (2.33)Remote 937 27 1405.00 1743.77 7199.6 3.07 (2.33)199411-199712 2773 48 1321.33 1460.87 8002.4 3.04 (2)199801-200112 3588 48 1436.47 3052.62 8868.1 3.67 (2.5)

35

Table IIRegressions of Relative Forecast Error on Analyst Characteristics

Fama and Macbeth (1973) cross-sectional regressions are run (yearly in Panel A, monthly in Panel B) from Novem-ber 1994 to December 2001. This table reports the time-series average of the estimated coefficients from regressionsof de-meaned absolute forecast error (DAFE) and proportional mean absolute forecast error (PMAFE) on a host ofanalyst characteristics. All variables are firm and fiscal-year mean adjusted (theD preceding each variable standsfor differenced or de-meaned). For example, the variableDAFE equals the difference between the absolute forecasterror for analysti for firm j in fiscal yeart and the mean absolute forecast error for firmj in fiscal yeart; abso-lute forecast error is equal to the absolute value of an analyst’s latest forecast, minus actual company earnings, as apercentage of stock price 12 months prior to the beginning of the fiscal year. The variablePMAFE equalsDAFEdivided by the mean absolute forecast error. The variableDAGE equals the log of one plus the age (in days) ofanalysti’s forecast,DBSIZEequals the log of one plus the number of analysts working for the same firm asi andsupplying forecasts during the same yeart, DSTARis a dummy variable equal to one if the analyst was ranked asan All-Star the previous October by Institutional Investor magazine,DHIGH is a dummy variable equal to one ifthe analyst works at a high-status firm (where status is measured by Institutional Investor magazine’s rankings ofbrokerage firms),DFEXPequals the log of one plus the number of days that analysti supplied a forecast for firmj,DPCTOPequals the percent of companies followed by analysti in yeart for which analysti’s last forecast is aboveactual earnings,DAFFIL is a dummy variable equal to one if the analyst works for the same brokerage firm thatserved as the lead underwriter for firmj ’s initial public offering within the past five years or for a secondary equityoffering by firm j within the past two years,DLOCAL is a dummy variable equal to one if the analyst is locatedwithin 100 kilometers of firmj in fiscal yeart, andDLOGDequals the log of one plus the distance (in kilometers) ofthe analysti from firm j in fiscal yeart; each variable is de-meaned as described above.t-statistics are in parentheses.

Panel A: Yearly RegressionsDep. Var. DAGE DBSIZE DSTAR DHIGH DFEXP DPCTOP DAFFIL DLOCAL DLOGDDAFE 0.2351 -0.0200 -0.0120 -0.0196 -0.0149 0.2567 -0.0609 -0.0197

(8.02) (-2.53) (-0.77) (-1.20) (-4.65) (6.57) (-2.80) (-2.13)PMAFE 38.3639 -3.9581 -3.8302 -5.8194 -1.9194 56.0157 -3.9867 -2.7652

(10.26) (-3.66) (-3.83) (-2.55) (-4.20) (7.92) (-2.47) (-2.28)DAFE 0.2350 -0.0201 -0.0119 -0.0198 -0.0149 0.2566 -0.0610 0.0025

(8.01) (-2.51) (-0.76) (-1.21) (-4.68) (6.56) (-2.78) (1.39)PMAFE 38.3497 -3.9959 -3.8217 -5.8506 -1.9221 56.0218 -3.9836 0.4649

(10.23) (-3.71) (-3.80) (-2.57) (-4.26) (7.91) (-2.46) (2.72)PMAFE 39.8310 -7.5650 -3.1704

(10.50) (-16.32) (-2.61)PMAFE 39.8150 -7.6174 0.5265

(10.47) (-16.03) (2.88)

Panel B: Monthly RegressionsDep. Var. DAGE DBSIZE DSTAR DHIGH DFEXP DPCTOP DAFFIL DLOCAL DLOGDPMAFE 43.8960 -3.0415 -5.6447 -5.4017 -2.1482 62.0247 -2.4224 -3.0436

(28.87) (-4.68) (-4.73) (-2.53) (-9.23) (16.82) (-0.36) (-2.31)PMAFE 43.8679 -3.1004 -5.6187 -5.3863 -2.1520 62.1661 -2.7410 0.5754

(28.80) (-4.74) (-4.73) (-2.51) (-9.25) (16.87) (-0.43) (2.33)

(Two-Year Ahead Forecasts)

PMAFE 70.5991 -1.6526 -0.1789 -1.1433 -0.5202 10.5647 -1.6196 -1.5152(20.10) (-3.67) (-0.27) (-1.50) (-3.99) (7.73) (-0.46) (-2.41)

PMAFE 70.5918 -1.6864 -0.1748 -1.1338 -0.5209 10.5731 -1.7182 0.3593(20.10) (-3.77) (-0.26) (-1.48) (-3.99) (7.72) (-0.50) (2.60)

36

Table IIIBias Versus Informativeness

Panel A reports the time series average of the estimated coefficients from yearly Fama and MacBeth (1973) re-gressions of de-meaned absolute forecast error (DAFE), proportional mean absolute forecast error (PMAFE), andrelative optimism (OPT, POPT) on a host of analyst characteristics. The dependent variablesDAFE andPMAFEand the independent variablesDAGE, DBSIZE, DSTAR, DHIGH, DFEXP, DAFFIL, DLOCAL, andDLOGD aredefined as in Table II; all variables are firm and fiscal-year mean adjusted (theD preceding each variable stands fordifferenced or de-meaned). The variableOPT equals the forecast error for analysti’s forecast of firm j for fiscalyeart (FEi, j,t ) minus the mean forecast error for firmj for fiscal yeart (FE j,t ); forecast error is equal to the analyst’slatest forecast, minus actual company earnings, as a percentage of stock price 12 months prior to the beginning of thefiscal year. The variablePOPT is simplyOPT scaled by the absolute value of (FE j,t ). Panel B reports the estimatedcoefficient, intercept, and R-squared values for Fama and MacBeth (1973) monthly cross-sectional regressions andpooled regressions of actual earnings on forecasted earnings; local analysts’ forecasts and non-local analysts’ fore-casts are grouped separately.t-statistics are in parentheses.

Panel A: Incorporating Relative Optimism into Table II RegressionsDep. Var. DAGE DBSIZE DSTAR DHIGH DFEXP (P)OPT DAFFIL DLOCALPMAFE 38.8645 -4.6365 -4.2789 -6.2852 -1.9307 1.5073 -2.6597 -2.7440

(10.19) (-4.54) (-4.28) (-2.71) (-4.16) (3.22) (-1.40) (-2.17)DAFE 0.1779 -0.01762 -0.0079 -0.0217 -0.0110 0.3648 -0.0705 -0.0153

(7.13) (-2.08) (-0.67) (-1.31) (-4.61) (5.90) (-4.91) (-1.66)POPT 57.2318 -10.2860 -3.2889 0.1960 -2.1762 5.5544 -1.2317

(8.76) (-6.14) (-0.75) (0.32) (-2.19) (1.60) (-0.85)OPT 0.1638 -0.0148 -0.0107 0.0003 -0.0097 0.0991 -0.0095

(10.91) (-1.78) (-1.02) (0.02) (-5.34) (1.62) (-1.11)

Panel B: Regressions of Actual Earnings on Forecasted Earnings(Ei, j,t = αi +βiFi, j,t + εi, j,t )

Fama-MacBeth PooledSAMPLE β α R2 β α R2

LOCAL= 1 0.9904 -0.0008 0.8589 0.9940 -0.0006 0.8752(155.26) (-3.20) (81.10) (206.07) (-4.52)

LOCAL= 0 0.9754 -0.0008 0.8215 0.9777 -0.0006 0.8342(206.07) (-4.36) (76.34) (631.68) (-7.71)

DIFF = (1−0) 0.0150 -0.0000 0.0374(1.95) (-0.06) (2.48)

37

Table IVLocal Regression Coefficients Sorted by Firm Characteristics

Fama and Macbeth (1973) cross-sectional regressions are run each year from 1994 to 2001 for sub-samples of TableII. This table reports only the coefficient estimates and associatedt-statistics (in parentheses) for theDLOCALvari-able in the Table II regression of de-meaned absolute forecast error (DAFE) on eight analyst characteristics. Stocksare classified as “Metro” if the firm’s headquarters is located in one the 20 most populated cities, and “Remote” ifthe firm’s headquarters is not located within 400 kilometers of any of the 20 most populated cities. The abbrevi-ation “Earn. Annc.” refers to quarterly earnings announcement dates.SMALLstocks are those below the medianmarket capitalization for all stocks in the sample, andLARGEstocks are those above the median market capitaliza-tion. GROWTHandVALUE stocks are sorted similarly but by book-to-market ratios;LOSERSandWINNERSaresorted similarly but by cumulative returns over the past year.t-statistics and the number of observations (n) are inparentheses.

SAMPLE Dep. Var. ALL SMALL LARGE GROWTH VALUE LOSERS WINNERSAll DAFE -0.0197 -0.0491 0.0017 -0.0109 -0.0335 -0.0287 -0.0150Stocks (-2.13) (-2.79) (0.31) (-1.20) (-1.74) (-2.69) (-1.25)(n=96,538)

Metro DAFE -0.0046 -0.0327 0.0021 -0.0047 -0.0092 -0.0020 -0.0109Stocks (-0.33) (-1.38) (0.23) (-0.30) (-0.39) (-0.13) (-0.88)(n=23,261)

Non-Metro DAFE -0.0266 -0.0579 0.0020 -0.0125 -0.0487 -0.0441 -0.0188Stocks (-2.12) (-2.44) (-0.35) (-0.82) (-2.22) (-3.48) (-1.03)(n=73,277)

Remote DAFE -0.0629 -0.0803 -0.0386 -0.0200 -0.1030 -0.0528 -0.0782Stocks (-2.68) (-3.19) (-1.03) (-1.03) (-2.21) (-2.11) (-3.07)(n=19,285)

Subperiod: DAFE -0.0136 -0.0433 0.0109 -0.0203 -0.0037 -0.0205 -0.0070199411-199712 (-0.75) (-1.22) (1.26) (-1.37) (-0.13) (-1.11) (-0.30)(n=40,616)

Subperiod: DAFE -0.0239 -0.0505 -0.0080 -0.0008 -0.0603 -0.0374 -0.0197199801-200112 (-3.50) (-3.65) (-2.65) (-0.07) (-3.59) (-2.77) (-2.58)(n=54,722)

Post-Reg FD: DAFE -0.0162 -0.0733 0.0339 0.0270 -0.0900 -0.0122 -0.0153200011-200112 (-0.32) (-1.18) (0.88) (0.41) (-1.87) (-0.20) (-0.33)(n=62,612)

Within 5 Days DAFE -0.0263 -0.0506 -0.0043 -0.0060 -0.0527 0.0236 -0.0438of Earn. Annc. (-0.94) (-1.03) (-0.56) (-0.47) (-1.05) (0.80) (-0.99)(n=25,915)

Non-Earn. DAFE -0.0249 -0.0627 -0.0012 -0.0028 -0.0549 -0.0381 -0.0148Annc. (-2.85) (-4.95) (-0.14) (-0.29) (-3.87) (-3.00) (-1.06)(n=61,886)

38

Table VLocal Regression Coefficients Sorted by Analyst Characteristics

Fama and Macbeth (1973) cross-sectional regressions are run each year from 1994 to 2001 for sub-samples ofTable II. This table reports only the coefficient estimates and associatedt-statistics (in parentheses) for theDLOCALvariable in the Table II regression of de-meaned absolute forecast error (DAFE) on eight analyst characteristics.SIGNAL= 1 if the analyst revises his forecast for firmj above his prior forecast for firmj and above the priorconsensus forecast for firmj; SIGNAL= −1 if the analyst revises his forecast for firmj below his prior forecastfirm j and below the prior consensus forecast for firmj. All other forecasts are classified asSIGNAL= 0. TheXNYCcolumn excludes all forecasts issued by analysts located within 100 kilometers of New York City, theSTARcolumn includes only those forecasts issued by All-Star analysts, theNON column excludes those forecasts issuedby All-Star analysts, theHIGH column includes only those forecasts issued by analysts working at a high-statusbrokerage house, and theLOW column excludes those forecasts issued by analysts at high-status brokerage houses.“Low-Coverage” analysts are those who fall below the 33rd percentile of number of stocks covered by an analyst in agiven year; “High-Coverage” are those above the 66th percentile.LOCMED is a dummy variable equal to one if theshare of local stocks in an analyst’s portfolio is above the median share of local stocks over all analysts’ portfolios.t-statistics and the number of observations (n) are in parentheses.

Panel A: By Forecast SignalSAMPLE Dep. Var. ALL XNYC STAR NON HIGH LOWAll DAFE -0.0197 -0.0250 -0.0462 -0.0116 -0.0714 -0.0192Forecasts (-2.13) (-2.76) (-1.64) (-1.12) (-1.17) (-2.07)(n=96,538)

SIGNAL= 1 DAFE -0.0207 -0.0192 0.0333 -0.0170 -0.0431 -0.0285Revisions (-1.15) (-0.91) (1.81) (-0.92) (-1.39) (-1.45)(n=33,910)

SIGNAL=−1 DAFE -0.0601 -0.0614 -0.1479 -0.0489 -0.1603 -0.0429Revisions (-3.31) (-2.90) (-1.53) (-2.42) (-0.64) (-3.23)(n=25,251)

SIGNAL= 0 DAFE -0.0081 -0.0318 -0.1742 0.0028 -0.0950 -0.0077Forecasts (-0.89) (-3.31) (-2.07) (0.31) (-1.76) (-0.94)(n=37,344)

39

Panel B: By Analyst PortfolioSAMPLE SUBGROUP Dep. Var. ALL XNYC STAR NON HIGH LOWLow-Coverage DAFE 0.0129 -0.0111 0.2851 0.0329 -0.0160 0.0216Analysts (0.45) (-0.28) (1.47) (1.02) (-0.17) (1.01)(n=11,107)

LOCMED=1 DAFE -0.0249 0.0174 0.1770 -0.014 -0.0458 0.00091(n=7,218) (-0.81) (0.21) (1.06) (-0.54) (-0.61) (0.28)

Mid-Coverage DAFE -0.0181 -0.0329 0.0472 -0.0299 -0.0116 -0.0251Analysts (-1.26) (-2.47) (0.55) (-1.24) (-1.09) (-1.96)(n=26,532)

LOCMED=1 DAFE -0.0400 -0.0377 0.0342 -0.0180 -0.0249 -0.0382(n=11,889) (-1.34) (-1.71) (0.52) (-0.38) (-0.56) (-1.86)

High-Coverage DAFE -0.0340 -0.0325 -0.1756 -0.0217 -0.1205 -0.0314Analysts (-2.78) (-2.93) (-3.10) (-2.09) (-1.29) (-2.13)(n=58,899)

LOCMED=1 DAFE -0.0512 -0.0626 -0.2070 -0.0471 -0.1616 -0.0572(n=18,734) (-2.67) (-2.16) (-2.70) (-2.61) (-1.23) (-2.42)

40

Table VIIndividual Analyst Portfolios and the Effect of Coverage Decisions

Panel A reports reports summary statistics broken down by analyst type. Analysts are classified each year as “Metro”and “Remote” each year as described in Table I. An analyst is classified as “Expert” in yeart if the analyst coversonly stocks falling within one 2-digit SIC code during yeart. An analyst is considered “Local” to firmj in yeart if the analyst is located within 100 kilometers of the headquarters of firmj in yeart. “No. of Obs.” equals thenumber of unique analyst-firm-year observations. “No of Firms Cov.” equals the average number of firms coveredeach year by each analyst type, and “No of Local Firms Cov.” equals the average number of local firms covered eachyear by each analyst type. “Pct. Expert” equals the percent of expert analysts across analyst types each year, and“No. of SIC Codes Cov.” equals the average number of 2-digit SIC codes covered by each analyst type each year.Panel B reports only the coefficient estimates and associatedt-statistics (in parentheses) for theDLOCALvariablein the Table II regression of de-meaned absolute forecast error (DAFE) on various analyst characteristics. The thirdcolumn adds the variableDTASK to this regression, the fourth column adds the variableDSPECinstead, and thefifth column adds bothDTASKandDSPECto the regression.DSPECequals the log of one plus the number offirms for which analysti supplied a forecast in yeart, andDPSECis a dummy variable equal to one if the analyst isan “Expert” as defined above; both variables are firm and fiscal-year mean adjusted. The sixth, seventh, and eighthcolumns of Panel B report estimates of theDLOCALvariable from the Table II regression, this time for subsamplesof firms broken down by advertising intensity (ADS/S).LOW (ADS/S) firms are those below the median level ofadvertising intensity for firms in the sample, andHIGH (ADS/S) are those above the median level.t-statistics andthe number of observations (n) are in parentheses.

Panel A: Summary Statistics for Individual Analyst PortfoliosAnalyst Type No. of Obs. No. of Firms No. of Local Pct. Expert No. of SIC Avg. Dist

Cov. Firms Cov. Analysts Codes Cov. to Firms (km)All Analysts 113517 12.37 2.01 32.66 2.13 1480.36Metro Analysts 88440 12.34 2.05 33.21 2.03 1567.50Remote Analysts 9689 11.10 2.22 31.63 2.40 1098.63Local Analysts 19185 11.87 4.17 32.51 2.14 28.79Expert Analysts 37072 10.43 1.67 100.00 1.00 1521.07

Panel B: Task Complexity, Specialization, and Investor DemandSAMPLE Dep. w/DTASK w/ DSPEC w/ DTASK NO LOW HIGH

Var. andDSPEC (ADS/S) (ADS/S) (ADS/S)All DAFE -0.0186 -0.0202 -0.0188 -0.0265 0.0377 -0.0510Stocks (-1.92) (-2.18) (-1.94) (-1.73) (1.63) (-3.25)(n=96,538)

Remote DAFE -0.0587 -0.0612 -0.0551 -0.0658 0.0075 -0.1184Stocks (-2.51) (-2.65) (-2.41) (-2.58) (0.07) (-1.51)(n=19,285)

41

Table VIIStock Price Impact of Forecast Revisions

Fama and Macbeth (1973) cross-sectional regressions are run every month from November 1994 to December 2001.This table reports the time-series average of the estimated coefficients from monthly regressions of percentage size-adjusted average daily returns around forecast revisions on firm size, the magnitude of the revision, the affiliationstatus of the analyst (AFFIL), and a measure of geographic proximity (LOCAL). Revisions are classified by signaltype, whereSIGNAL= 1 if the analyst revises his forecast for firmj above his prior forecast for firmj and abovethe prior consensus forecast for firmj; SIGNAL= −1 if the analyst revises his forecast for firmj below his priorforecast firm j and below the prior consensus forecast for firmj. Stocks are classified as “Metro” if the firm’sheadquarters is located in one the 20 most populated cities. The variableLOCALi is a dummy variable equal toone if the analyst is located within 100 kilometers of firmj. The variableAFFILi is a dummy variable equal toone if the analyst works for the same brokerage firm that served as the lead underwriter for firmj ’s initial publicoffering within the past five years or for a secondary equity offering by firmj within the past two years. The variableAARj,a,b is the percentage size-adjusted average daily return for firmj from event daya to b. The variableSUFi, j,0

is the standardized unexpected forecast for analysti, which is defined as the change in the forecast divided by thecross-sectional standard deviation of outstanding forecasts for firmj. The variableLNME− j,−30 is the log of oneplus the market capitalization of firmj one month prior to the revision.t-statistics and the number of observations(n) are in parentheses.

Panel A:AARj,−1,1 = β0 +β1LNMEj,−30+β2SUFi, j,0 +β3AFFILi +β4LOCALi + εSAMPLE Signal Type β0 β1 β2 β3 β4

All SIGNAL= 1 0.8469 -0.0539 -0.0769 0.0735 0.0946Stocks (n=86,127) (7.38) (-4.41) (-0.75) (0.68) (2.38)

SIGNAL=−1 -0.5643 0.0492 0.9781 -0.5550 -0.1281(n=63,214) (-2.56) (1.75) (5.54) (-3.67) (-2.46)

Metro SIGNAL= 1 0.8836 -0.0586 -0.1172 0.1320 0.0395Stocks (n=23,097) (4.68) (-2.59) (-0.43) (0.70) (0.55)

SIGNAL=−1 -0.7551 0.0816 0.5343 -0.4660 -0.1346(n=13,894) (-2.70) (2.39) (3.26) (-1.99) (-1.61)

Non-Metro SIGNAL= 1 0.8368 -0.0538 -0.0494 0.0403 0.1019Stocks (n=63,030) (7.26) (-4.17) (-0.34) (0.34) (1.95)

SIGNAL=−1 -0.4893 0.0345 0.9927 -0.3640 -0.1638(n=49,320) (-1.90) (1.08) (5.72) (-2.26) (-2.54)

Panel B:AARj,2,64 = β0 +β1LNMEj,−30+β2SUFi, j,0 +β3AFFILi +β4LOCALi + εSAMPLE Signal Type β0 β1 β2 β3 β4

All SIGNAL= 1 0.1205 -0.0091 -0.0475 -0.0102 0.0020Stocks (n=86,127) (4.88) (-3.15) (-2.80) (-0.60) (0.33)

SIGNAL=−1 -0.0536 -0.0012 0.0455 -0.0005 -0.0188(n=63,214) (1.74) (-0.35) (2.53) (-0.02) (-1.09)

42

Panel C:AARj,−1,1 = β0 +β1LNMEj,−30+β2SUFi, j,0 +β3AFFILi +β4LOCALi + εSAMPLE Signal Type β0 β1 β2 β3 β4

Subperiod: SIGNAL= 1 0.3888 -0.0196 0.0354 0.4245 0.0566199411-199712 (n=34,067) (3.43) (-1.32) (0.67) (3.47) (1.99)

SIGNAL=−1 -0.4389 0.0320 0.4648 -0.4584 -0.0580(n=27,635) (-3.65) (2.37) (5.97) (-2.11) (-1.19)

Subperiod: SIGNAL= 1 1.2076 -0.0811 -0.1620 -0.1897 0.1256199801-200112 (n=51,867) (7.23) (-4.63) (-0.92) (-1.20) (1.87)

SIGNAL=−1 -0.6677 0.0629 1.369 -0.6648 -0.1802(n=35,382) (-1.75) (1.29) (4.63) (-3.22) (-2.15)

Post-Reg FD: SIGNAL= 1 1.6342 -0.1121 -0.1149 -0.1653 0.1361200011-200112 (n=5,506) (3.42) (-2.35) (-0.20) (-0.41) (0.67)

SIGNAL=−1 -1.8525 0.1976 1.4686 -1.0540 -0.4210(n=5,419) (-1.54) (1.27) (1.50) (-2.35) (-1.93)

Panel D:AARj,−1,1 = β0 +β1LNMEj,−30+β2SUFi, j,0 +β3AFFILi +β4LOCALi + εSAMPLE Signal Type β0 β1 β2 β3 β4

Excl. NYC SIGNAL= 1 0.6949 -0.0344 0.0322 0.2245 0.2277Forecasts (n=37,007) (4.94) (-2.22) (0.23) (1.10) (2.35)

SIGNAL=−1 -0.3176 0.01411 0.8140 -0.3294 -0.0779(n=26,179) (-2.19) (0.73) (6.55) (-1.21) (-0.82)

All-Star SIGNAL= 1 1.0315 -0.0704 -0.0764 -0.3586 0.2157Forecasts (n=15,919) (5.67) (-3.29) (-0.53) (-1.84) (2.09)

SIGNAL=−1 -0.6176 0.0454 0.9746 -0.5244 -0.0551(n=12,070) (-3.01) (1.86) (6.13) (-1.95) (-0.43)

Non-All-Star SIGNAL= 1 0.8368 -0.0533 -0.0860 0.2068 0.0592Forecasts (n=70,208) (6.37) (-4.08) (-0.69) (1.64) (1.31)

SIGNAL=−1 -0.6211 0.0616 1.0595 -0.4102 -0.1740(n=51,144) (-2.22) (1.62) (4.35) (-2.41) (-2.88)

High-Status SIGNAL= 1 1.2245 -0.0927 -0.0321 -0.1690 0.3173Forecasts (n=26,169) (9.00) (-5.73) (-0.35) (-1.20) (3.81)

SIGNAL=−1 -0.4255 0.0248 0.9827 -0.6068 -0.1443(n=20,067) (-2.36) (1.22) (8.10) (-2.41) (-1.36)

Low-Status SIGNAL= 1 0.7667 -0.0458 -0.1014 0.2629 0.0387Forecasts (n=59,958) (6.26) (-3.61) (-0.80) (1.64) (0.83)

SIGNAL=−1 -0.6231 0.0616 0.9441 -0.3707 -0.1437(n=43,147) (-2.79) (2.14) (5.37) (-1.72) (-2.86)

High Advertising SIGNAL= 1 2.3208 -0.2110 0.4549 0.0887 0.1284Forecasts (n=11,743) (1.93) (-1.69) (1.56) (0.81) (1.47)

SIGNAL=−1 -0.8234 0.0804 0.9164 -0.4191 -0.1393(n=6,964) (-2.13) (1.62) (4.20) (-1.53) (-0.66)

43

Table VIIIStock Price Impact of Affiliated Analyst Recommendations

Fama and Macbeth (1973) cross-sectional regressions are run every month from November 1994 to December 2001.This table reports the time-series average of the estimated coefficients from monthly regressions of percentage size-adjusted average daily returns around stock recommendations on a series of dummy variables. In Panels A, B, andC, the last four independent variables are equal to one if the recommendation is a strong buy (SBj ), buy (B j ), hold(H j ), or sell (Sj ). In Panel A, the first three independent variables (SBj,A, B j,A, andH j,A) are equal to one if therecommendation is a strong buy, buy, hold, or sell, and is issued by an analyst who works for the same brokeragefirm that served as the lead underwriter for the firm’s most recent secondary equity offering (within the past twoyears), and equal to zero otherwise. Thusβ1, β2, andβ3 capture the mean incremental returns associated withaffiliated analysts’ recommendations relative to the same recommendations by non-affiliated analysts. In Panel B,the first three independent variables (SBj,A,L, B j,A,L, andH j,A,L) are equal to one if the recommendation is a strongbuy, buy, hold, or sell, and is issued by a lead underwriter analyst who is also local (local analysts reside within 100kilometers of the firm’s headquarters). The first three independent variables (SBj,A,NL, B j,A,NL, andH j,A,NL) in PanelC are defined similarly; the subscriptNL refers to a lead underwriter analyst who isnot a local analyst. The variableAARj,a,b is the percentage size-adjusted average daily return from event daya to b. t-statistics and the number ofobservations (n) are in parentheses.

Panel A: Incremental Returns Associated With All AffiliatesAARj,−a,b = β1SBj,A +β2B j,A +β3H j,A +β4SBj +β5B j +β6H j +β7Sj + ε

Dep. Var. β1 β2 β3 β4 β5 β6 β7

Announcement Period: AARj,−1,1 0.62211 -0.6862 -1.7602 0.7198 -0.0390 -1.1553 -1.0672(n=86,824) (1.26) (-2.18) (-3.78) (16.11) (-1.04) (-14.23) (-7.37)

Pre-Release Period:AARj,−22,−2 -0.0741 -0.1071 -0.0751 0.0463 0.0335 -0.0689 -0.0557(n=86,824) (-1.30) (-1.87) (-1.17) (4.99) (2.70) (-4.70) (-2.06)

Panel B: Incremental Returns Associated With Local AffiliatesAARj,a,b = β1SBj,A,L +β2B j,A,L +β3H j,A,L +β4SBj +β5B j +β6H j +β7Sj + ε

Dep. Var. β1 β2 β3 β4 β5 β6 β7

Announcement Period: AARj,−1,1 0.3322 -0.4214 -0.2463 0.7202 -0.0438 -1.1635 -1.0786(n=86,824) (1.96) (-1.75) (-0.78) (16.18) (-1.16) (-14.26) (-7.38)

Pre-Release Period:AARj,−22,−2 0.0019 0.0050 -0.021 0.0457 0.0327 -0.0683 -0.0556(n=86,824) (0.04) (0.07) (-0.69) (4.91) (2.63) (-4.74) (-2.07)

Panel C: Incremental Returns Associated With Distant AffiliatesAARj,a,b = β1SBj,A,NL +β2B j,A,NL +β3H j,A,NL +β4SBj +β5B j +β6H j +β7Sj + ε

Dep. Var. β1 β2 β3 β4 β5 β6 β7

Announcement Period: AARj,−1,1 0.0826 -0.6966 -1.738 0.7211 -0.0402 -1.1562 -1.0672(n=86,824) (0.46) (-2.10) (-3.87) (16.12) (-1.07) (-14.27) (-7.37)

Pre-Release Period:AARj,−22,−2 -0.0673 -0.09106 -0.0676 0.0464 0.0335 -0.0680 -0.0557(n=86,824) (-1.16) (-1.75) (-0.98) (4.98) (2.70) (-4.71) (-2.06)

44

Table IXThe Affiliation Bias in Analyst Stock Recommendations

Panel A reports differences in mean recommendation codes (1=“Strong Buy”, 5=“Strong Sell”), where recommen-dations are matched in pairs using a (maximum) 60 day window between recommendation dates. The variableAFFIL is a dummy variable equal to one if the analyst works for the same brokerage firm that served as the leadunderwriter for the firm’s most recent secondary equity offering (SEO), and zero otherwise;NONAFF=1 wheneverAFFIL=0, and vice-versa. The variableLOCAFF is a dummy variable equal to one ifLOCAL=1 andAFFIL=1,and zero otherwise; the variableDISTAFF is a dummy variable equal to one ifLOCAL=0 andAFFIL=1, and zerootherwise. In theALL column, the affiliated recommendation nearest to the SEO date is chosen, while in theALL90,ALL120, andALL250 columns, the affiliated recommendation must be within 90, 120, or 250 days of the SEO date,respectively. TheMET column includes only those stocks headquartered in one the 20 most populated cities, whiletheNMET column excludes those metropolitan stocks. Panel B presents summary statistics for the latter two sam-ples of matched recommendations.META refers to analysts located in a metro area, whileMETF refers to stockslocated in a metro area; all other variables are described in Table V.t-statistics and the number of matched pairs (n)are in parentheses.

Panel A: Matched Pair Mean DifferencesALL MET NMET ALL90 ALL120 ALL250

AFFIL 1.7247 1.6098 1.7500 1.5238 1.5484 1.5912NONAFF 1.8947 2.0000 1.8702 1.7551 1.7742 1.8343DIFF -0.1700 -0.3902 -0.1202 -0.2313 -0.2258 -0.2431

(-2.45) (-2.16) (-1.61) (-2.81) (-2.78) (-3.11)(n=247) (n=41) (n=206) (n=147) (n=155) (n=181)

LOCAFF 1.7105 1.5000 1.7500 1.5000 1.5217 1.6000NONAFF 1.7105 1.6667 1.7188 1.5909 1.5652 1.6000DIFF 0.0000 -0.1667 0.0312 -0.0909 -0.0435 0.0000

(0.00) (-0.42) (0.18) (-0.48) (-0.23) (0.00)(n=38) (n=6) (n=32) (n=22) (n=23) (n=25)

DISTAFF 1.7395 1.6486 1.7611 1.5280 1.5530 1.5897NONAFF 1.9209 2.0270 1.88944 1.7840 1.8106 1.8718DIFF -0.1814 -0.3784 -0.1333 -0.2560 -0.2576 -0.2821

(-2.41) (-1.96) (-1.64) (-2.81) (-2.87) (-3.31)(n=209) (n=35) (n=174) (n=125) (n=132) (n=156)

Panel B: Matched Pair Summary StatisticsPct.HIGH Pct.STAR Pct.NYC Pct.META Pct.METF

LOCAFF 34.21 15.79 31.58 76.32 15.79NONAFF 14.29 5.26 31.58 76.32 15.79(n=38)

DISTAFF 55.35 30.23 68.84 91.63 16.74NONAFF 19.90 9.77 42.33 69.30 16.74(n=209)

45

Table XWhy are Recommendations By Local Affiliates Unbiased?

Panel A reports summary statistics for the entire recommendation sample. The columnALL includes all recommen-dations, theAFFIL column includes only affiliated recommendations, etc.; the variablesAFFIL, LOCAFF, andDISTAFF are defined as in Table VIII.FIRST further limits the sample to affiliated recommendations associatedwith first-time SEOs, whileREPEATincludes all other affiliated recommendations. Panel B reports differences inmean recommendation codes, as in Panel A of Table VIII. The variableX is associated with the analyst character-istic in the designated column. For example,AFFILX=1 refers to affiliated recommendations where the affiliatedanalyst is also in categoryX; the first row of theHIGH column thus reports the mean recommendation code foraffiliated analysts who also work at high-status firms. All affiliated recommendations are matched to non-affiliatedrecommendations as described in Table VIII.t-statistics and the number of matched pairs (n) are in parentheses.

Panel A: Recommendation Sample BreakdownALL AFFIL AFFIL AFFIL LOCAFF DISTAFF

(FIRST) (REPEAT)Pct.HIGH 21.09 46.62 46.13 46.60 30.56 49.43

Pct.LOCAL 16.83 14.90 15.81 14.36 100.00 0.00

Panel B: Matched Pair Mean Differences(Nearest Affiliated Rec. to SEO) (Affiliated Rec. Within 90 Days of SEO)

X = [ HIGH LOW NYC XNYC HIGH LOW NYC XNYCAFFILX=1 1.7333 1.7339 1.7212 1.7872 1.5190 1.5294 1.5319 1.5094NONAFF 1.9852 1.8145 1.9152 1.8404 1.8481 1.6471 1.8404 1.6048DIFF -0.2519 -0.0808 -0.1939 -0.0532 -0.3291 -0.1176 -0.3085 -0.0943

(-2.64) (-0.84) (-2.29) (-0.47) (-2.89) (-0.99) (-3.04) (-0.67)(n=136) (n=111) (n=152) (n=95) (n=79) (n=68) (n=94) (n=53)

LOCAFFX=1 1.7692 1.6923 1.7500 1.6923 1.4286 1.5333 1.7500 1.3571NONAFF 1.7692 1.6923 1.8333 1.6538 1.5714 1.6000 1.6250 1.5714DIFF 0.0000 0.0000 -0.083 -0.0385 -0.1429 -0.0667 0.1250 -0.2143

(0.00) (0.00) (-0.28) (-0.02) (-0.50) (-0.26) (0.40) (-0.89)(n=12) (n=26) (n=12) (n=26) (n=7) (n=15) (n=8) (n=14)

DISTAFFX=1 1.7419 1.7549 1.7124 1.8406 1.5278 1.5283 1.5116 1.5641NONAFF 2.0081 1.8333 1.9281 1.8986 1.8750 1.6604 1.8605 1.6154DIFF -0.2661 -0.0784 -0.2157 -0.0580 -0.3472 -0.1321 -0.3488 -0.0513

(-2.66) (-0.73) (-2.44) (-0.42) (-2.85) (-0.97) (-3.26) (-0.30)(n=124) (n=85) (n=140) (n=69) (n=72) (n=53) (n=86) (n=39)

X = 1 1.7669 1.8229 1.6169 1.7514X = 0 1.8496 1.8192 1.8182 1.7946DIFF -0.0827 0.0040 -0.2012 -0.0432

(-1.24) (0.06) (-2.48) (-0.58)(n=266) (n=206) (n=155) (n=185)

46